{
  "schema_version": "1.0",
  "dataset_updated_at": "2026-10-18T05:54:49.495524+00:00",
  "count": 270,
  "models": [
    {
      "slug": "deepseek-v41-flash-1017",
      "source_file": "2026-10-17-deepseek-v4.1-flash-1017.json",
      "release_date": "2026-10-17",
      "date": "October 17, 2026",
      "name": "DeepSeek-V4.1-Flash-1017",
      "developer": "DeepSeek AI (China)",
      "parameters": "552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "DeepSeek describes DeepSeek-V4.1-Flash-1017 as a minor update to the September DeepSeek-V4.1-Flash checkpoint that delivers substantial improvements across reasoning, coding, agentic, and visual benchmarks. The company says the update also reduces task execution time and output length, allowing users to complete tasks faster and at lower cost while advancing the performance-cost Pareto frontier across multiple agent harnesses. Publisher-reported results versus the September checkpoint include HLE without tools 42.6 versus 36.8, Terminal-Bench 4.0 42.3 versus 31.2, DeepSWE v1.1 77.0 versus 74.2, ProgramBench 29.1 versus 20.3, NL2Repo-Bench 68.2 versus 65.4, Automation-Bench 60.5 versus 54.8, Agents' Last Exam 35.4 versus 31.8, and Apex Bench 72.1 versus 64.6. Tool-assisted Chartography, BabyVision, and ZeroBench-main rise to 83.0, 92.5, and 51.0 respectively. Terminal-Bench 4.0 uses the Claude Code harness, while DeepSWE uses Mini-SWE-V2. The update retains the V4.1 architecture and 1,048,576-token context window: a 40-layer Causal Encoder-Decoder with a 20-layer causal encoder and 20-layer decoder, 552B backbone parameters, 196B of Engram conditional-memory parameters, 8B activated during prefill, and 16B activated during decoding. Each MoE layer uses one shared expert and 384 routed experts with six selected per token. The architecture combines Compressed Sparse Attention 2, FP4 main-KV caching, SWA Bounded Replay, Single-Pass mHC residual mixing, two sparsely accessed Engram modules, and integrated DSpark speculative decoding. A 32-layer DeepSeek-ViT encoder provides native image understanding with up to 1,024 image tokens. DeepSeek does not disclose the update's additional training data, token budget, compute, or exact post-training recipe. Official Hugging Face and ModelScope repositories provide the MIT-licensed checkpoint in 48 mixed-precision safetensors shards totaling approximately 510.3 GB. The release does not include a Jinja chat template; DeepSeek instead supplies reference encoding code and the maintained deepseek-recipe toolkit for multimodal conversations, thinking, numeric reasoning effort, tool calls, and Chat Completions or Responses API conversion. Recommended settings are temperature 1.0, top-p 1.0, a 1M-token context window, and at least 256K maximum output tokens.",
      "paper_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/config.json",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/tree/main/encoding",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/tree/main/inference",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf",
        "https://github.com/deepseek-ai/deepseek-recipe"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v41-flash-1017/",
      "record_updated_at": "2026-10-18T05:54:49.495524+00:00"
    },
    {
      "slug": "nvidia-nemotron-labs-dory-10-4b",
      "source_file": "2026-10-16-nvidia-nemotron-labs-dory1.0-4b.json",
      "release_date": "2026-10-16",
      "date": "October 16, 2026",
      "name": "NVIDIA Nemotron Labs Dory 1.0 4B",
      "developer": "NVIDIA Corporation (USA)",
      "parameters": "3.98B physical parameters (looped dense model)",
      "context_window": "262,144 tokens (256K)",
      "license": "OpenMDW License Agreement, version 1.1",
      "key_features": "Small text-only reasoning and agent model built around a looped, or latent-recurrent, normalized-GPT architecture. Dory divides its network into eight input layers, a shared 20-layer recurrent block, and eight output layers, reinjecting the input into each recurrence. The same checkpoint can execute the recurrent block one, two, or four times without retraining; NVIDIA recommends two loops as the default balance between compute and accuracy. Its attention pattern interleaves sliding-window and full-attention layers. Per-loop KV caching retains separate caches for each recurrence, while the optional last-loop mode overwrites them so KV-cache memory remains at approximately the one-loop footprint regardless of loop count. The BF16 repository contains exactly 3,975,229,440 parameters in one 7.95 GB safetensors file, with official MXFP8, NVFP4, and Q5_K_M GGUF releases also available. NVIDIA pretrained Dory on 6T tokens, using native NVFP4 for approximately the first 4.5T tokens and BF16 for the remainder, then performed a 300B-token context-extension phase, supervised fine-tuning, and asynchronous multi-environment GRPO across mathematics, coding, science, instruction following, structured output, multi-step tool use, and multi-turn conversation. The released chat template supports reasoning and tool calling. NVIDIA reports SWE-bench Verified scores of 50.3, 59.5, and 61.8 and Terminal-Bench 2.1 scores of 12.5, 18.1, and 22.2 at one, two, and four loops respectively; these are publisher evaluations under NVIDIA's stated harness. Native 256K inference is documented through vLLM, but Dory support remains under review upstream in vLLM and llama.cpp at release time.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-MXFP8",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-NVFP4",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-GGUF",
        "https://developer.nvidia.com/topics/ai/nemotron",
        "https://huggingface.co/datasets/nvidia/nemotron-pre-training-datasets",
        "https://huggingface.co/datasets/nvidia/nemotron-post-training-v3",
        "https://github.com/vllm-project/vllm/pull/60149",
        "https://github.com/ggml-org/llama.cpp/pull/30306",
        "https://openmdw.ai/license/1-1/"
      ],
      "page_url": "https://nlp.com.ai/models/nvidia-nemotron-labs-dory-10-4b/",
      "record_updated_at": "2026-10-17T02:52:36.084118+00:00"
    },
    {
      "slug": "beam-10",
      "source_file": "2026-10-16-beam-1.0.json",
      "release_date": "2026-10-16",
      "date": "October 16, 2026",
      "name": "Beam 1.0",
      "developer": "Reflection AI",
      "parameters": "501.3B total / 23.4B active (sparse MoE)",
      "context_window": "1,048,576 tokens",
      "license": "Apache 2.0",
      "key_features": "Reflection's first open-weight model, released as an English-predominant, text-only instruction model for software engineering, reasoning, tool use, browsing, and long-horizon agentic workloads. Beam 1.0 is a sparse mixture-of-experts Transformer with 501.3B total parameters and 23.4B activated per token. Its 52-layer language backbone begins with three dense layers followed by 49 MoE layers, each containing 256 routed experts and selecting eight experts per token. The released configuration uses a 5,120-dimensional hidden state, 20,480-dimensional dense feed-forward layers, 2,560-dimensional expert feed-forward layers, 48 query heads, eight key/value heads, 128-dimensional attention heads, an untied 151,936-token vocabulary, and a modified Harmony prompt format. Gated grouped-query attention combines mostly 512-token sliding-window layers with periodic full-attention layers. The native configuration and tokenizer specify 1,048,576 positions, and Reflection lists a June 30, 2026 knowledge cutoff. Beam was pretrained on 23.8 trillion tokens from curated public web data, source code, technical documentation, educational and scientific material, and commercially licensed datasets. Mid-training adds curated code repositories, long-form documents, tool-use trajectories, and long-horizon tasks. Reinforcement learning spans coding, agentic, search, tool-use, and STEM environments generated primarily through synthetic pipelines and supplemented by proprietary vendor and open-source data. Reflection reports more than 100 million high-compute RL rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. Across training, evaluation, and ablations, the company reports approximately 1.3 billion sandbox instances; the peak hour of the final RL run reached 883,006 allocatable vCPUs across 23 GCP and AWS clusters. Reflection positions Beam as reaching comparable solution quality to leading open models while consuming approximately three times fewer inference-time reasoning tokens. Final developer-reported BF16 scores include 43.3 on DeepSWE v1.1, 76.5 on SWE-bench Pro v2-Hard, 66.1 on SWE-bench Pro v1, 80.1 on Terminal-Bench 2.1, 78.3 on SWE-bench Multilingual, 81.6 on SWE-bench Verified, 97.7 on AIME 2026, 36.0 on Humanity's Last Exam without tools, 51.1 with tools, 90.6 on GPQA Diamond, 88.3 on LiveCodeBench, 80.8 on MCP Atlas, 36.1 on tau3-bench Banking, 76.8 on BrowseComp with context management, 80.5 on DeepSearchQA with context management, 67.5 on LongBench v2, and 79.6 on IFBench. Reflection reports broadly similar results from the compressed checkpoints, including several cases where FP8 or NVFP4 slightly exceeds BF16. The release includes 188 BF16 safetensors shards totaling approximately 1.003 TB, 188 FP8 shards totaling approximately 503.3 GB, and 53 NVFP4 shards totaling approximately 293.6 GB. A separate 1.93B-parameter DFlash2 speculative-decoding model uses five layers, a 4,096-token sliding window, and hidden states drawn from five Beam layers to draft blocks of tokens which Beam then verifies. On four NVIDIA B300 GPUs with Beam NVFP4, Reflection reports mean speculative acceptance lengths of 3.93 to 5.60 tokens and greedy-decoding throughput gains ranging from 1.72 to 3.03 times across GSM8K, MATH-500, HumanEval, MBPP, and MT-Bench. Day-zero deployment support includes Transformers, vLLM, SGLang, NVIDIA Dynamo, TokenSpeed, Unsloth, and Axolotl. Reflection additionally released the Mirror CLI coding and agent harness, and reports testing Beam with Pi, OpenCode, and Hermes. The launch includes customization integrations with NeMo RL, AutoModel, Megatron, Miles, TRL, SkyRL, Axolotl, and OpenEnv, with self-distillation recipes announced for a subsequent release. Reflection's Iris Safety Evaluations report an 88.9% harmful-request refusal-or-pivot rate and an 81.5% harmless-request compliance rate for BF16. The Apache-2.0 release explicitly permits commercial and research use, fine-tuning, parameter-efficient adaptation, quantization, merging, distillation, and using model outputs to train other models. Default evaluation sampling uses temperature 0.7 and top-p 0.99.",
      "paper_url": "https://reflection.ai/beam-safety-report",
      "links": [
        "https://reflection.ai/blog/beam-is-available",
        "https://reflection.ai/blog/introducing-beam",
        "https://huggingface.co/collections/reflection/beam-10",
        "https://huggingface.co/reflection/Beam-1.0-501B-A23B",
        "https://huggingface.co/reflection/Beam-1.0-501B-A23B-FP8",
        "https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4",
        "https://huggingface.co/reflection/Beam-1.0-501B-A23B-DFlash2",
        "https://reflection.ai/beam-model-card",
        "https://reflection.ai/beam-safety-report",
        "https://developers.reflection.ai/introduction",
        "https://github.com/reflection-oss/mirror"
      ],
      "page_url": "https://nlp.com.ai/models/beam-10/",
      "record_updated_at": "2026-10-16T13:31:26.173882+00:00"
    },
    {
      "slug": "ternary-bonsai-2-2b-4b-9b",
      "source_file": "2026-10-15-bonsai-2-small.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "Ternary Bonsai 2 2B / 4B / 9B",
      "developer": "Prism ML, Inc. (United States)",
      "parameters": "1.88B / 4.21B / 8.95B language model + 0.33B / 0.33B / 0.46B vision tower",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Grouped release of three compact second-generation ternary vision-language models derived from Qwen3.5-2B, Qwen3.5-4B, and Qwen3.5-9B with their base architectures unchanged. The 2B model contains approximately 1.88B language-model parameters across 24 blocks plus a 0.33B, 24-block vision tower; the 4B contains approximately 4.21B language-model parameters across 32 blocks plus the same 0.33B vision tower; and the 9B contains approximately 8.95B language-model parameters across 32 blocks plus a 0.46B, 27-block vision tower. All three retain Qwen3.5's 262,144-token context window, native reasoning and non-thinking modes, image understanding, and structured tool-calling capabilities. Their hybrid-attention language backbones use a repeating 3:1 ratio of linear-attention to full-attention layers with SwiGLU MLPs, RoPE, and RMSNorm. Language-model matrix weights are represented as ternary {-1, 0, +1} values with FP16 group-wise scales and a group size of 128. A fixed blockwise Walsh-Hadamard rotation is folded into the stored weights, while normalization weights and the linear-attention state path remain in F32. The smallest PTQ1_0 GGUF language-model packs occupy approximately 0.43 GB, 0.95 GB, and 2.00 GB, versus 3.76 GB, 8.41 GB, and 17.91 GB for the corresponding BF16 language models, making them approximately 8.8-9.0 times smaller. Prism ML also provides faster-to-unpack PQ2_0 packs, Q2_0 packs compatible with recent upstream llama.cpp, MLX two-bit packages for Apple Silicon, Q8_0 and BF16 vision projectors, and optional MTP and DFlash2 speculative-decoding components for the 4B and 9B variants. Across Prism ML's evaluation suite, reported overall averages are 61.81 versus 63.95 for the Qwen3.5-2B baseline, 73.36 versus 77.53 at 4B, and 77.02 versus 80.96 at 9B. Relative to their same-size BF16 baselines, the ternary models retain 98.2-107.8% of math performance, 90.0-95.5% of agentic and tool-calling performance, 100.1-115.0% of coding performance, 100.0-121.5% of instruction-following performance, and 89.0-92.5% of vision performance. The 9B reports 82.5 on AIME 2026, 89.02 on HumanEval+, 71.81 on BFCL v3, and 78.02 on BFCL v4 Mobile. Reported batch-one decode throughput using PTQ1_0 reaches approximately 603.6, 368.6, and 248.6 tokens per second for the 2B, 4B, and 9B models on an RTX 4090. MLX throughput on an Apple M5 Pro reaches approximately 276.3, 145.2, and 91.7 tokens per second, while an iPhone 17 Pro Max reaches approximately 89.9, 39.8, and 22.3 tokens per second. Prism ML reports 3.6-4.2 times lower generation energy per token than BF16 on an RTX 4090 and 1.7-2.0 times lower energy on an M5 Pro. The models were trained using Google TPU v5 hardware. Deployment uses Prism ML's custom llama.cpp kernels for the PTQ1_0 and PQ2_0 formats or its MLX-compatible runtime, with official examples covering OpenAI-compatible serving, multimodal prompting, local tool use, and speculative decoding.",
      "paper_url": "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-small-whitepaper.pdf",
      "links": [
        "https://prismml.com/news/prismml-expands-bonsai-2-family-with-2b-4b-and-9b-models-for-on-device-ai",
        "https://huggingface.co/collections/prism-ml/bonsai-2-small",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-2B-gguf",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-2B-mlx-2bit",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-4B-gguf",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-4B-mlx-2bit",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-9B-gguf",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-9B-mlx-2bit",
        "https://github.com/PrismML-Eng/Bonsai-demo",
        "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-small-whitepaper.pdf",
        "https://github.com/PrismML-Eng/llama.cpp",
        "https://github.com/PrismML-Eng/mlx",
        "https://huggingface.co/Qwen/Qwen3.5-2B",
        "https://huggingface.co/Qwen/Qwen3.5-4B",
        "https://huggingface.co/Qwen/Qwen3.5-9B"
      ],
      "page_url": "https://nlp.com.ai/models/ternary-bonsai-2-2b-4b-9b/",
      "record_updated_at": "2026-10-15T20:42:28.434864+00:00"
    },
    {
      "slug": "muse-spark-open",
      "source_file": "2026-10-15-muse-spark-open.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "Muse Spark Open",
      "developer": "Meta Superintelligence Labs (USA)",
      "parameters": "816B total / 42B active MoE (approximately 822.0B serialized BF16 parameters)",
      "context_window": "1M tokens",
      "license": "Muse Model License",
      "key_features": "Meta's unnumbered open-weight Muse Spark release for long-horizon agentic work, coding, tool use, multimodal reasoning, synthetic-data generation, and LLM-as-a-judge evaluation. Muse Spark Open is a 62-layer sparse Mixture-of-Experts Transformer with an 8,192-dimensional hidden state, gated attention with QK normalization, 128 query heads, 16 key-value heads, a 202,048-token vocabulary, and 256 routed experts selecting eight per token. Meta advertises 816B total and 42B active parameters, while Hugging Face inventories approximately 822.0B serialized parameters in the BF16 checkpoint. A separate approximately 1.8B-parameter, 50-layer perception encoder with a 1,536-dimensional hidden state, 16 attention heads, 14-by-14 patches, sparse attention, learned absolute positions, and 2D RoPE supports text, images, documents, audio, and video as input with text output. The model supports a 1M-token context window, more than 100 training languages, multi-step planning, sequential tool invocation, failure recovery, coding across large repositories, web and computer use, subagent coordination, and low, medium, high, and xhigh reasoning effort. Meta describes this as a distinct open release of Muse Spark with additional safety post-training intended to improve robustness when deployed without the system-level guardrails surrounding hosted Muse checkpoints. Official BF16 and NVFP4 checkpoints are available through manually gated Hugging Face repositories; Meta states that BF16 deployment requires approximately 2TB of GPU memory and NVFP4 approximately 700GB. Supported launch runtimes include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth, with verified NVIDIA, AMD, Intel, and Apple deployments; Meta and Apple demonstrated local operation across connected M5 Ultra Mac Studios. Meta reports 61.8 on JobBench, 89.2 on MCP-Atlas, 80.4 on OSWorld Verified, 26.1 on HLE-Diamond without tools, 58.5 on SciCode, 89.0 on GPQA Diamond, 60 on DeepSWE 1.1, 76.4 on Terminal-Bench 2.1, 50.0 on ZeroBench with tools, 73.6 on SimpleVQA with tools, and 61.5 on VideoMMEv2 at xhigh reasoning. The custom Muse Model License allows use, modification, redistribution, derivatives, and commercial use subject to its usage policy and attribution requirements, while qualifying Model-as-a-Service businesses whose affiliated revenue exceeds $20 million must notify Meta and enter a separate written agreement.",
      "paper_url": "https://research.meta.ai/static/muse-spark-open-evaluation-methodology",
      "links": [
        "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B",
        "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B-NVFP4",
        "https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF",
        "https://research.meta.ai/blog/muse-spark-open",
        "https://dev.meta.ai/resources/blog/build-with-muse-spark-open",
        "https://dev.meta.ai/docs/muse-spark-open/get-the-model",
        "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md"
      ],
      "page_url": "https://nlp.com.ai/models/muse-spark-open/",
      "record_updated_at": "2026-10-15T17:21:44.893546+00:00"
    },
    {
      "slug": "comet-60b",
      "source_file": "2026-10-15-comet-60b.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "Comet 60B",
      "developer": "Multiverse Computing (Spain)",
      "parameters": "61.1B total / 12.3B active (marketed as 60B total / 12B active MoE)",
      "context_window": "262,144 tokens evaluated (configuration permits 1,048,576 positions)",
      "license": "NVIDIA Nemotron Open Model License",
      "key_features": "Open-weight multimodal reasoning and agent model produced by compressing NVIDIA Nemotron 3.5 Super VL with Multiverse Computing's quantum-inspired CompactifAI framework. Comet reduces the parent model from 124.5B total parameters to 61.1B while retaining approximately the same active compute at 12.3B parameters per token. It preserves the 88-layer hybrid Mamba-2, Transformer, latent Mixture-of-Experts, attention, RADIO ViT-H/16 vision, and Multi-Token Prediction architecture, while reducing the number of routed experts per MoE layer from 512 to 256 and continuing to activate 22 experts plus one shared expert. Comet accepts text, images, and video and produces text, with configurable thinking, non-thinking, tool calling, multimodal document understanding, coding, and agentic operation; audio is not supported. After tensor-network profiling and structural compression, Multiverse healed the model through token-level knowledge distillation from Nemotron 3.5 Super using single-turn reasoning and code, multi-turn terminal and tool-use trajectories, images, and video. A final GRPO-style stage targeted multi-turn interaction, tool use, hallucination reduction, reasoning, and verbosity. The family includes a 122.3GB BF16 checkpoint requiring two H200s for the full evaluated 256K context, a 66.3GB FP8 checkpoint that fits on one H200, and a 40.5GB mixed NVFP4/FP8/BF16 checkpoint for Blackwell hardware. The released configuration permits 1M positions, but Multiverse reports evaluation only through 262,144 tokens. Multiverse reports BF16 scores of 88.64 on GPQA Diamond, 70.00 on SWE-bench Verified, 50.93 on SWE-bench Pro, 80.79 on tau2-Bench, 33.18 on AutomationBench-AA, 67.40 on MMMU-Pro, and 73.26 on Video-MME with subtitles. The FP8 checkpoint scores 71.00 on SWE-bench Verified, 51.48 on SWE-bench Pro, 81.01 on tau2-Bench, and 30.93 on BabyVision. In Multiverse's H200 serving tests, BF16 on four GPUs delivered 20% more chat and 23% more agentic output tokens per second per GPU than Nemotron 3.5 Super, while FP8 on two GPUs delivered 87% and 78% more respectively.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/MultiverseComputingCAI/Comet_60B_A12B_BF16",
        "https://huggingface.co/MultiverseComputingCAI/Comet_60B_A12B_FP8",
        "https://huggingface.co/MultiverseComputingCAI/Comet_60B_A12B_NVFP4",
        "https://www.globenewswire.com/news-release/2026/10/15/3381915/0/en/multiverse-computing-launches-comet-60b-in-collaboration-with-nvidia-multimodal-capabilities-at-half-the-parameters.html",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/comet-60b/",
      "record_updated_at": "2026-10-15T16:34:13.330780+00:00"
    },
    {
      "slug": "nvidia-nemotron-35-bolt-74b-a9b",
      "source_file": "2026-10-15-nvidia-nemotron-3.5-bolt-74b-a9b.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "NVIDIA Nemotron 3.5 Bolt 74B-A9B",
      "developer": "NVIDIA Corporation (USA)",
      "parameters": "74B total / 9B active (LatentMoE)",
      "context_window": "262,144 tokens (256K)",
      "license": "OpenMDW License Agreement, version 1.1",
      "key_features": "Open-weight vision-language reasoning and agent model compressed from NVIDIA Nemotron 3.5 Super VL 120B-A12B for local deployment. Bolt combines a hybrid Mamba-2, Transformer, Mixture-of-Experts, and attention language backbone with a C-RADIO v4-H vision encoder and MLP projector, accepting text, images, and video while producing text. NVIDIA used its Star Elastic router-guided architecture-search and knowledge-distillation method to reduce the parent model from 120B total / 12B active parameters to 74B total / 9B active, lowering expert top-K from 22 to 9. The native IQ2 release uses quantization-aware training to pack the model into approximately 24GB, allowing it to run alongside local agent applications; full BF16 and IQ2 GGUF checkpoints are also available. Training included compression and recovery over approximately 80B tokens, vision-adapter recovery, vision-heavy supervised fine-tuning, checkpoint merging, RLHF, and end-to-end GRPO in MCP and software-engineering environments. The model supports switchable thinking and non-thinking modes, tool calling, web browsing, terminal use, document and screen understanding, and English, Spanish, French, German, Italian, Japanese, and programming languages. A companion LiLiCorr draft model uses DFlash to propose seven-token blocks for speculative decoding and reports a 4.30-token average accepted length on SPEED-Bench. NVIDIA reports IQ2 scores of 50.98 on Terminal-Bench 2.1, 86.43 on GPQA Diamond, 66.75 on BrowseComp, 52.20 on MCP-Atlas, and 50.86 on vision WebArena; the BF16 checkpoint scores 53.93, 88.83, 70.00, 55.60, and 50.25 respectively.",
      "paper_url": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-Technical-Report.pdf",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF",
        "https://research.nvidia.com/labs/nemotron/Nemotron-3.5-Bolt-24GB/",
        "https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/",
        "https://huggingface.co/collections/nvidia/nvidia-nemotron-v3",
        "https://huggingface.co/datasets/nvidia/nemotron-pre-training-datasets",
        "https://huggingface.co/datasets/nvidia/nemotron-post-training-v3",
        "https://openmdw.ai/license/1-1/"
      ],
      "page_url": "https://nlp.com.ai/models/nvidia-nemotron-35-bolt-74b-a9b/",
      "record_updated_at": "2026-10-15T16:09:07.408873+00:00"
    },
    {
      "slug": "gpt-usenet-6",
      "source_file": "2026-10-15-gpt-usenet-6.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "GPT-USENET-6",
      "developer": "HDTenEightyP",
      "parameters": "0.4B dense (421,772,288 serialized parameters)",
      "context_window": "16,384 tokens",
      "license": "MIT License",
      "key_features": "Small English base language model using the Transformers GPT2LMHeadModel architecture with the cl100k_base tokenizer associated with GPT-4-era models. Its configuration contains 24 transformer layers, 16 attention heads, 1,024-dimensional embeddings, and a 100,277-token vocabulary. The developer reports training it for 24 hours on a Google Colab G4 using approximately 30GB of USENET posts, web scrapes, and stories, reaching a training loss of 2.5916 and validation loss of 2.7335. The released checkpoint uses FP32 weights and supports a 16,384-token context window. It is a raw text-completion model rather than an instruction-following assistant, and its model card explicitly recommends supervised fine-tuning before practical use.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/HDTenEightyP/GPT-USENET-6"
      ],
      "page_url": "https://nlp.com.ai/models/gpt-usenet-6/",
      "record_updated_at": "2026-10-15T15:11:03.278950+00:00"
    },
    {
      "slug": "nvidia-nemotron-35-super-vl-120b-a12b",
      "source_file": "2026-10-15-nvidia-nemotron-3.5-super-vl-120b-a12b.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "NVIDIA Nemotron 3.5 Super VL 120B-A12B",
      "developer": "NVIDIA Corporation (USA)",
      "parameters": "120B total / 12B active (multimodal MoE)",
      "context_window": "1,048,576 tokens (1M)",
      "license": "OpenMDW License Agreement, version 1.1",
      "key_features": "Open-weight vision-language reasoning and agent model combining interleaved Mamba-2, sparse Mixture-of-Experts, and attention layers with a C-RADIO vision encoder and vision-language connector. It accepts text, images, and video and produces text, supporting reasoning, coding, tool use, document understanding, visual grounding, OCR, and video understanding. The model has 512 routed experts per MoE layer, activates 12B of its 120B parameters per token, and includes a shared Multi-Token Prediction head for native speculative decoding. Its chat template supports switchable thinking and non-thinking operation. NVIDIA extended the language backbone through software-engineering mid-training, 1M-context training, multimodal training, and post-training using reasoning and agent trajectories; the release also links its pretraining and post-training datasets and reproducible NeMo Gym evaluation recipes. Official BF16 reference weights target customization, domain adaptation, and full-precision evaluation and can be deployed on four H100s or two GB200s, while the companion NVFP4 checkpoint targets more efficient serving. NVIDIA reports 89.52 on GPQA Diamond, 88.22 on LiveCodeBench v6, 71.10 on SWE-bench Verified, 51.78 on SWE-bench Pro, 66.70 on Terminal-Bench 2.1, 68.12 on English OCRBench v2, 80.82 on CharXiv reasoning, and 51.14 on OSWorld-Verified.",
      "paper_url": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Super-and-Lightning-Technical-Report.pdf",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4",
        "https://huggingface.co/datasets/nvidia/nemotron-pre-training-datasets",
        "https://huggingface.co/datasets/nvidia/nemotron-post-training-v3",
        "https://github.com/NVIDIA-NeMo/Gym/blob/main/nemotron_recipes/super-3.5/reproducibility.md",
        "https://openmdw.ai/license/1-1/"
      ],
      "page_url": "https://nlp.com.ai/models/nvidia-nemotron-35-super-vl-120b-a12b/",
      "record_updated_at": "2026-10-15T15:11:03.271863+00:00"
    },
    {
      "slug": "jais-3",
      "source_file": "2026-10-15-jais-3.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "Jais 3",
      "developer": "Institute of Foundation Models, MBZUAI / Inception42 (UAE)",
      "parameters": "0.9B, 3.7B, 7B, and 32B dense; 36B total / 4B active (MoE + MoVA); 375B total / 23B active (MoE)",
      "context_window": "512K tokens (128K for the 0.9B model)",
      "license": "Apache License 2.0",
      "key_features": "Six-model open-weight family focused on Arabic-native reasoning, instruction following, and agentic work while retaining English and general reasoning capabilities from the underlying K2 Horizon models. Jais 3 covers Modern Standard Arabic, Emirati and Egyptian dialects, Arabizi, Arabic poetry and cultural knowledge, and is trained to plan multi-step tasks, call tools and MCP servers, search the web, run terminal commands, and edit Arabic documents, presentations, and spreadsheets. The family comprises four dense checkpoints at 0.9B, 3.7B, 7B, and 32B; a 36B-A4B sparse model combining Mixture-of-Experts layers with Mixture-of-Values Attention; and a 375B-A23B flagship MoE. Every model except the 0.9B supports a 512K-token context window; the 0.9B supports 128K. IFM released final weights, intermediate checkpoints from each training stage, configs, recipes, logs, sample training data, audited Arabic benchmarks, and evaluation code. Training includes Arabic continued pretraining, supervised agent data, Final Token Preference Optimization for reducing repetitive reasoning loops, and verifiable-reward reinforcement learning across instruction following, Arabic grammatical analysis, mathematics, diacritization, and code. IFM reports AA-LCR scores of 65 for Jais 3 7B versus 49 for Gemma 4 12B, and 75 for Jais 3 32B versus 66 for Gemma 4 31B. Optional rank-128 Uno adapters are available for the four dense models, converting them to discrete-diffusion generation with speculative sampling; the Uno release reports an average 2.15x H100 speedup while targeting the original model distribution. BF16, FP8, GGUF, and MLX variants are available for supported sizes.",
      "paper_url": "",
      "links": [
        "https://ifm.ai/blog/jais3/",
        "https://huggingface.co/collections/IFM/jais-3",
        "https://huggingface.co/inception42/Jais-3-0.9B",
        "https://huggingface.co/inception42/Jais-3-3.7B",
        "https://huggingface.co/inception42/Jais-3-7B",
        "https://huggingface.co/inception42/Jais-3-32B",
        "https://huggingface.co/inception42/Jais-3-MoVA-36B-A4B",
        "https://huggingface.co/inception42/Jais-3-375B-A23B",
        "https://huggingface.co/papers/2609.04010",
        "https://github.com/ifm-ai/uno"
      ],
      "page_url": "https://nlp.com.ai/models/jais-3/",
      "record_updated_at": "2026-10-15T15:11:03.267670+00:00"
    },
    {
      "slug": "step-5-preview",
      "source_file": "2026-10-15-step-5-preview.json",
      "release_date": "2026-10-15",
      "date": "October 15, 2026",
      "name": "Step 5 Preview",
      "developer": "StepFun (China)",
      "parameters": "600B total / 27B active (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "StepFun Model License 1.0 (current official distribution; September 20 leaked snapshot carried Apache-2.0 metadata)",
      "key_features": "Official open-weight preview of StepFun's flagship model for software engineering, long-horizon agents, professional knowledge work, and finance. Step 5 Preview is a native multimodal sparse Mixture-of-Experts model marketed as 600B total and 27B activated parameters; Hugging Face inventories approximately 604.34B serialized parameters. Its narrow-and-deep language backbone has 92 transformer blocks, a 4,096-dimensional hidden state, 64 attention heads in four groups, 352 routed experts selecting eight per token, three initial dense layers, a 128,896-token vocabulary, and three attached multi-token-prediction layers. Attention follows a repeating 3:1 pattern of 512-token Gated Sliding Window Attention and full-attention layers using StepFun's Sparse GQA, which combines sparse indexing with block-wise token merging for million-token operation. The released configuration natively specifies 1,048,576 positions. A 47-layer, 1,536-wide vision encoder with 16 heads and 14-by-14 patches provides visual input; StepFun describes text, image, and video support, although the concise GitHub specification lists text and image input while the hosted API documentation additionally documents video. The official release includes a blockwise FP8 E4M3 checkpoint with dynamic activation quantization and a BF16 checkpoint. Each contains 24 main language-model shards plus vision and MTP artifacts; all safetensors total approximately 650.3 GB for FP8 and 1.215 TB for BF16. StepFun's technical report describes mixed-FP8 training with FP32 hidden states, a role-based agentic data-synthesis system, bitwise training-inference alignment across tensor-parallel sizes, RL trace replay for routed experts and top-p candidate sets, context compaction trained end to end, and token- and turn-level credit assignment for very long trajectories. StepFun reports an end-to-end rollout sample-drop rate below 1%. At High reasoning effort, reported results include 67.7 on DeepSWE v1.1, 80.5 on ProgramBench, 49.0 on StepCodeBench, 33.3 on Terminal-Bench v4, 29.5 on Agents' Last Exam CLI, 66.4 on FrontierFinance, and 83.3 on DRACO. StepFun's official September 20 FP8 repository history labeled the accidentally exposed snapshot `apache-2.0`, although it included no license text. A surviving BF16 fork contains safetensors byte-identical to the October 15 official release. StepFun subsequently replaced the Apache metadata with StepFun Model License 1.0. Whether the original metadata created an enforceable Apache grant for leaked copies remains unresolved. Official deployment paths cover vLLM, SGLang, and standalone Transformers on B300 and H200 systems with tensor, data, and expert parallelism plus EAGLE speculative decoding. The hosted API supports low, medium, and high reasoning, tool calling, prompt caching, JSON Mode, JSON Schema, and a documented 64K maximum output; the GitHub model specification separately claims output up to 1M tokens.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/stepfun-ai/Step-5-Preview",
        "https://huggingface.co/stepfun-ai/Step-5-Preview-BF16",
        "https://modelscope.cn/models/stepfun-ai/Step-5-Preview",
        "https://modelscope.cn/models/stepfun-ai/Step-5-Preview-BF16",
        "https://github.com/stepfun-ai/Step-5-Preview",
        "https://www.stepfun.com/step-5-preview",
        "https://platform.stepfun.ai/blog/inside-step-5-preview",
        "https://platform.stepfun.ai/docs/en/guides/models/step-5-preview",
        "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/9953fd1981bdc98aebc9a2538bed95816e469cac/README.md",
        "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/main/LICENSE.txt"
      ],
      "page_url": "https://nlp.com.ai/models/step-5-preview/",
      "record_updated_at": "2026-10-15T09:09:39.956078+00:00"
    },
    {
      "slug": "beellm-20",
      "source_file": "2026-10-14-beellm-2.0.json",
      "release_date": "2026-10-14",
      "date": "October 14, 2026",
      "name": "BeeLLM 2.0",
      "developer": "schizophyllume (skizy)",
      "parameters": "1.1M and 11,965,760 weights (dense Llama architecture)",
      "context_window": "4,096 tokens",
      "license": "Apache 2.0",
      "key_features": "Second-generation release of an intentionally tiny family of open-weight bee language models designed to buzz, express moods, perform actions, remember how they are treated within the conversation, and pollinate flowers. BeeLLM 2.0 contains two dense Llama-architecture models distributed directly as F32 GGUF files for llama.cpp: BeeLLM-2.0-1M-Mood, with approximately 1.1 million weights across four transformer layers and a 1,024-token vocabulary, and BeeLLM-2.0-12M-Mood, with exactly 11,965,760 weights across eight layers and a 2,048-token vocabulary. Both use a 4,096-token context window and a custom chat template with <|user|> and <|bee|> roles. The 1M checkpoint occupies approximately 4.49 MB and the 12M checkpoint approximately 47.91 MB. Both were retrained on approximately 3,000 new conversations relative to the first-generation Mood models. Prism-like compression is not involved: these are genuinely very small, full-F32 language models. Responses begin with a mood-and-intensity tag such as [happy:high], [angry:low], or [confused:high], followed by buzzes, short words, and physical actions. The models can react differently to flowers, affection, swatting, bug spray, confinement, apologies, and attempts at immediate lovebombing. The 12M lineage knows that her name is Scintilla, can develop grudges against objects and actions, may file formal complaints against jars, and very rarely stings the user and dies before potentially reviving several turns later. BeeLLM 2.0 gives both variants a richer voice and a favorite flower; the 1M version is friendlier toward gentle messages, while the 12M version is additionally described as kinder to other bots and particularly fond of orange flowers. The broader repository also contains the original 30P and 42P tiny-buzz models, Chaos variants with shouting and stinging vocabulary, and first-generation 1M and 12M Mood checkpoints. All models run anywhere supported by llama.cpp and are small enough to embed in web applications. The official interactive browser demonstration presents the model as a bee with feelings that remembers how the user treats her. No conventional benchmark suite or technical paper is provided. The model card lists its language as buzz, its intended uses as buzzing and pollinating flowers, and its primary limitation as the possibility that it may sting.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/schizophyllume/BeeLLM",
        "https://huggingface.co/schizophyllume/BeeLLM/commit/f469ed610f716ae438c04c2bd04921fed7eabd55",
        "https://bee.pleroma.nexus/",
        "https://github.com/ggml-org/llama.cpp"
      ],
      "page_url": "https://nlp.com.ai/models/beellm-20/",
      "record_updated_at": "2026-10-15T21:45:35.955012+00:00"
    },
    {
      "slug": "ling-31-flash",
      "source_file": "2026-10-13-ling-3.1-flash.json",
      "release_date": "2026-10-13",
      "date": "October 13, 2026",
      "name": "Ling-3.1-flash",
      "developer": "inclusionAI (China)",
      "parameters": "560B total / 25B active (hybrid-linear MoE)",
      "context_window": "Up to 1M tokens advertised (262,144-token checkpoint configuration)",
      "license": "MIT",
      "key_features": "MIT-licensed open-weight upgrade to Ling-3.0-flash and native hybrid-reasoning text-generation model aimed at general agents, search, office productivity, software engineering, healthcare, finance, and materials science. Its hybrid-linear Mixture-of-Experts backbone has 72 transformer layers arranged as 63 Kimi Delta Attention layers and nine Gated Multi-head Latent Attention layers in a repeating 7:1 pattern, with four initial dense layers, 512 routed experts selecting eight per token, one shared expert, 64 attention heads, a 4,096-dimensional hidden state, and a 157,184-token vocabulary. inclusionAI reports 560B total parameters and 25B activated per token; Hugging Face's safetensors metadata inventories approximately 572.1B serialized parameters, while the linked SGLang cookbook estimates roughly 17B active from its expert accounting. The released checkpoint uses blockwise FP8 E4M3 weights with dynamic activation quantization and includes one NEXTN multi-token-prediction layer for speculative decoding. Its 54 downloadable safetensors shards total approximately 573.8 GB. The model card advertises a context window up to 1M tokens, although the checkpoint configuration and verified SGLang deployment documentation specify 262,144 positions, so the advertised million-token operating regime is not encoded in the published configuration. Thinking is enabled by default but can be disabled, and the supplied chat template preserves reasoning across turns while supporting structured tool calls. inclusionAI reports 1,673 Elo on GDPVal-AA V2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional. The release also introduces LingCode, a Pi-derived harness for multi-hour coding tasks, hundreds of tool calls, and repeated context compaction, but inclusionAI has not yet open-sourced that harness. Official deployment guidance covers SGLang and vLLM, including tensor and expert parallelism, prefix caching, structured tool and reasoning parsers, and NEXTN speculative decoding; the linked SGLang recipe reports 251.2 output tokens per second at batch size one on eight H200 GPUs.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-3.1-flash",
        "https://modelscope.cn/models/inclusionAI/Ling-3.1-flash",
        "https://huggingface.co/inclusionAI/Ling-3.1-flash/blob/main/config.json",
        "https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.1-flash",
        "https://huggingface.co/inclusionAI/AI-Transparency/resolve/main/Ling-3.1_TDS-Summary.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/ling-31-flash/",
      "record_updated_at": "2026-10-14T18:00:53.273821+00:00"
    },
    {
      "slug": "lfm25-24b-a2b",
      "source_file": "2026-10-13-lfm2.5-24b-a2b.json",
      "release_date": "2026-10-13",
      "date": "October 13, 2026",
      "name": "LFM2.5-24B-A2B",
      "developer": "Liquid AI (USA)",
      "parameters": "24.0B total / 2.45B active (MoE)",
      "context_window": "131,072 tokens",
      "license": "LFM Open License v1.0",
      "key_features": "Same-day open-weight release of base and post-trained text-only checkpoints for Liquid AI's largest and final LFM2.5 model. Its sparse hybrid architecture contains 40 layers comprising 30 double-gated short-convolution blocks and 10 grouped-query-attention blocks, with 64 experts per MoE block and top-4 routing. The release doubles the predecessor's vocabulary from 65K to 128K tokens, extends context from 32,768 to 131,072 tokens, supports 16 languages, and reports a 36T-token training budget: approximately 33T pretraining tokens followed by 2T tokens of curated 32K-context mid-training and 300B tokens at 131K context. The post-trained checkpoint is a reasoning-only model whose chat template always begins responses with an explicit chain of thought. Its five-stage post-training pipeline combines supervised fine-tuning, teacher specialization, non-agentic and agentic Multi-Domain On-Policy Distillation, and agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi, and OpenCode. Liquid AI reports 48.09 on LiveClaw, 27.69 on AutomationBench, 65.01 on BFCLv4, 80.71 on ToolSandbox, 96.75 on IFEval, and 51.00 on SWE-bench Verified using OpenCode. Reported deployment performance reaches 184 tokens per second on an Apple M5 Max, 114 tokens per second on an AMD Ryzen AI Max+ 395, and more than 21K aggregate output tokens per second on one NVIDIA B300 at high concurrency. Official deployment formats include native BF16, FP8, GGUF, ONNX, and MLX checkpoints with support for Transformers, vLLM, SGLang, llama.cpp, and Apple Silicon.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B",
        "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-Base",
        "https://www.liquid.ai/blog/lfm2-5-24b-a2b",
        "https://docs.liquid.ai/lfm/models/lfm25-24b-a2b"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-24b-a2b/",
      "record_updated_at": "2026-10-13T18:05:58.466373+00:00"
    },
    {
      "slug": "minicpm-v-47-35b-a3b-1b",
      "source_file": "2026-10-12-minicpm-v-4.7.json",
      "release_date": "2026-10-12",
      "date": "October 12, 2026",
      "name": "MiniCPM-V 4.7 (35B-A3B, 1B)",
      "developer": "OpenBMB / ModelBest (China)",
      "parameters": "35B-A3B: 35.21B stored / approximately 3B active (MoE); 1B: 1.55B stored (1.3B-class dense)",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Same-day open-weight release of the MiniCPM-V 4.7 multimodal family in 35B-A3B and 1B variants, supporting text, single-image, multi-image, document, and video inputs with textual responses. The flagship checkpoint contains exactly 35,212,875,824 BF16 parameters and uses the MiniCPM-V family's first Mixture-of-Experts language backbone, derived from Qwen3.6. Its 40-layer hybrid stack repeats three linear-attention layers followed by one full-attention layer, contains 256 routed experts with eight selected per token plus a shared expert, and includes a one-layer multi-token-prediction module. The compact variant contains exactly 1,554,476,272 BF16 parameters, uses a dense 24-layer Qwen3.5-derived language backbone, and follows the same three-to-one linear/full-attention pattern. Both variants pair their language model with a 27-layer SigLIP2-derived vision encoder, support mixed 4x and 16x visual-token compression, and use canvas M-RoPE to preserve two-dimensional relationships among image slices and temporal order among video frames. OpenBMB positions both models on a new multimodal performance-efficiency Pareto frontier. Its instruct-mode evaluation reports 91.72 on MMBench, 80.22 on MMMU, 91.20 on OCRBench, 94.24 on OmniDocBench v1.6, 76.40 on Video-MME without subtitles, and 68.90 on LongVideoBench for the 35B-A3B model at 4x visual compression. Reported language scores for that checkpoint include 90.21 on CMMLU, 76.46 on AIME26@avg16, 56.57 on LiveCodeBench v6, and 84.65 on tau2-Bench. The 1B model reports 83.17 on MMBench, 86.10 on OCRBench, 68.05 on ScreenSpot, and 54.50 on LongVideoBench, with OpenBMB claiming that it substantially exceeds Qwen3.5-0.8B in both overall score and inference efficiency. For local MoE deployment, OpenBMB introduces Lookahead Expert Scheduling, which predicts cross-layer routing only for expert prefetch and scheduling while preserving the model's original token-level routing. Combined with fine-grained layer-wise quantization and DSpark speculative sampling, OpenBMB reports 13-plus decoding tokens per second using approximately 3 GB of working memory on an iPhone 17 Pro Max. The release supports iOS, Android, HarmonyOS, Intel, AMD, Arm, Rockchip, and Houmo hardware, with deployment through Transformers, SGLang, vLLM, llama.cpp, Ollama, and multiple official BF16, FP8, int4, AWQ, GPTQ, GGUF, and DSpark artifacts. OpenBMB also released Self-Replay Fine-Tuning for extending model capabilities while reducing forgetting.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/collections/openbmb/minicpm-v-47",
        "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B",
        "https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-35B-A3B",
        "https://huggingface.co/openbmb/MiniCPM-V-4.7-1B",
        "https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-1B",
        "https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/minicpm_v4dot7_en.md",
        "https://huggingface.co/docs/transformers/main/model_doc/minicpmv4_7",
        "https://ollama.com/openbmb/minicpm-v4.7"
      ],
      "page_url": "https://nlp.com.ai/models/minicpm-v-47-35b-a3b-1b/",
      "record_updated_at": "2026-10-13T08:41:32.300355+00:00"
    },
    {
      "slug": "colgrep-agent-2b",
      "source_file": "2026-10-12-colgrep-agent-2b.json",
      "release_date": "2026-10-12",
      "date": "October 12, 2026",
      "name": "ColGrep-Agent-2B",
      "developer": "LightOn (France)",
      "parameters": "2.52B dense (2,516,756,480 BF16 parameters; marketed as 2B)",
      "context_window": "131,072 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight compact code-localization agent built on MiniCPM5-2B and designed to identify the files and line ranges relevant to GitHub issues, bug reports, feature requests, and natural-language questions about software repositories. ColGrep-Agent-2B uses a dense 42-layer Llama-style causal architecture with approximately 2.52B parameters and a 131,072-token context window. Rather than generating patches itself, it explores repositories through three tools: ColGrep semantic code search, a read-only terminal for inspecting candidate files, and a finish tool that submits relevant locations in `path/to/file:START-END` form. The supplied agent harness gives it a ten-turn exploration budget and is intended to let larger coding agents delegate repository search to a small locally hosted specialist. LightOn trained the model with a short supervised cold start using trajectories generated and filtered by Qwen3.5-27B, followed by 500 GRPO reinforcement-learning steps rewarding accurate file- and line-level localization while penalizing unnecessary exploration. Training environments were constructed from SWE-rebench, SWE-Fixer-Train-110K, and additional multilingual repositories paired with issues and bug-fix commits. On LightOn's five-seed code-localization evaluation, the model reports 76.18 File F1 and 19.92 Line F1 on SWE-bench Lite, versus 66.72 and 9.62 for its MiniCPM5-2B base model. On Multi-SWE-bench Flash it reports 62.57 File F1 and 15.93 Line F1, versus 52.01 and 11.31 for the base model. These measurements use 300 instances per benchmark per seed, a ten-turn limit, temperature 0.6, and no reasoning for comparison models; they evaluate localization rather than end-to-end issue resolution. The official release includes approximately 5.04 GB of BF16 safetensors and a separate GGUF repository with multiple quantizations. ColGrep's installer defaults to the approximately 2.5 GB Q8_0 build and runs the agent locally without requiring Python, an external inference server, or an API key.",
      "paper_url": "https://arxiv.org/abs/2610.05099",
      "links": [
        "https://huggingface.co/lightonai/colgrep-agent-2B",
        "https://huggingface.co/lightonai/colgrep-agent-2B-GGUF",
        "https://huggingface.co/blog/lightonai/colgrep-agent/",
        "https://github.com/lightonai/next-plaid",
        "https://arxiv.org/abs/2610.05099",
        "https://www.lighton.ai/about"
      ],
      "page_url": "https://nlp.com.ai/models/colgrep-agent-2b/",
      "record_updated_at": "2026-10-12T14:59:25.303705+00:00"
    },
    {
      "slug": "mrchatterbox-27b-chat",
      "source_file": "2026-10-06-mr-chatterbox-2.7b-chat.json",
      "release_date": "2026-10-06",
      "date": "October 6, 2026",
      "name": "Mr.Chatterbox-2.7B-Chat",
      "developer": "Trip Venturella",
      "parameters": "2.7B dense (2,705,820,160 serialized parameters)",
      "context_window": "2,048 tokens (fine-tuned with a 1,024-token maximum sequence length)",
      "license": "CC BY-NC-SA 4.0",
      "key_features": "Conversational version of Mr.Chatterbox-2.7B, a language model trained from scratch on British material published between 1800 and 1899 rather than a modern model prompted to imitate Victorian prose. The 2.7B base model was pretrained on 54.1B tokens. Chat fine-tuning used 19,939 conversations expanded into 46,353 assistant-turn examples; the developer reports that 97.2% of the assistant replies are verbatim 19th-century text, with the remainder consisting of lightly edited period text or model-generated rejection-fine-tuning examples. Every assistant response begins internally with one of 20 learned conversational-act tokens, such as answer_fact, explain, compose, admit_ignorance, clarify, or rebuke, allowing the type of response to be steered before generation. Fine-tuning ran for two epochs and 18,172 steps on one NVIDIA B200 in approximately 80 minutes. An LLM-judged 1,356-prompt behavioral evaluation gave this checkpoint 0.751 for responsiveness, behavioral fit, and period register, compared with 0.655 for the original 345M Mr. Chatterbox. The project presents the model as an experimental historical conversation artifact rather than a factual assistant and warns that its unfiltered period training corpus can reproduce offensive Victorian attitudes. The current weights use CC BY-NC-SA 4.0 because portions of the source corpus carry noncommercial and share-alike restrictions; the repository was initially published under MIT before being relicensed on October 7.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/tventurella/Mr.Chatterbox-2.7B-Chat",
        "https://huggingface.co/tventurella/Mr.Chatterbox-2.7B",
        "https://huggingface.co/datasets/tventurella/Mr.Chatterbox-SFT",
        "https://github.com/tripv/mr.chatterbox"
      ],
      "page_url": "https://nlp.com.ai/models/mrchatterbox-27b-chat/",
      "record_updated_at": "2026-10-15T15:11:03.275443+00:00"
    },
    {
      "slug": "kolibri-1",
      "source_file": "2026-10-03-kolibri-1.json",
      "release_date": "2026-10-03",
      "date": "October 3, 2026",
      "name": "Kolibri-1",
      "developer": "Aleph Alpha (Germany)",
      "parameters": "78.1B total / 3.46B active (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight English-German reasoning and tool-use model designed by Aleph Alpha for sovereign deployment, regulated industries, public administration, industrial applications, and aerospace. Kolibri-1 is a 50-layer decoder-only causal Transformer with 78,103,074,560 total parameters and 3,457,573,120 active parameters per token, activating approximately 4.4% of its parameters. Its hybrid-attention architecture contains 40 sliding-window-attention blocks with a 512-token window and ten full-attention blocks arranged at a 4:1 ratio. The model has a width of 2,560, 48 query heads, four KV heads, and 128-dimensional attention heads. Every layer contains a sparse MoE comprising 384 routed experts, of which six are selected per token, plus one shared expert. The architecture uses sandwich RMS normalization, grouped-query attention, Muon optimization, and Exact Quantile Balancing for expert routing. Kolibri was trained on approximately 24 trillion tokens across three stages: 20T tokens of pre-training at 16K context, 3.44T tokens of mid-training at 64K, and 201B tokens of long-context extension at 256K. The pre-training corpus was approximately 62.5% English, 23.9% German, and 13.6% code, including more than 2T German tokens curated from the web or generated synthetically. Its 128,000-token UniBPE vocabulary was developed to improve German morphological alignment and compression while retaining competitive English efficiency. Post-training combines bilingual supervised fine-tuning with reinforcement learning across more than 1.2 million internally curated tasks spanning reasoning, instruction following, code, retrieval, and tool use. The model supports low, medium, and high reasoning-effort levels, as well as a non-thinking mode, and exposes Hermes-style structured tool calls through Aleph Alpha's open-source vLLM plugin. Although its final long-context training used sequences up to 262,144 tokens, Aleph Alpha validated quality and serving efficiency at 1,048,576 tokens and recommends contexts at or below 262,144 for complex or throughput-sensitive workloads. Aleph Alpha reports scores of 96.9 on AIME 2025, 96.0 on AIME 2026, 84.3 on GPQA Diamond, 85.9 on LiveCodeBench v6, 92.7 on HumanEval+, 61.4 on BFCL v4, 64.5 on LongBench Pro, 69.9 on tau2-bench Retail, and 38.1 on tau3-bench Banking. Pre-training used 768 NVIDIA B200 GPUs for 21 days, followed by five days of mid-training and thirteen hours of long-context training. The released FP8 checkpoint uses float8_e4m3fn weights in 128-by-128 blocks with dynamically quantized activations while retaining embeddings, the language-model head, normalization layers, and MoE routers in BF16; its 32 weight shards total approximately 78.84 GB. A separate 32-shard BF16 checkpoint totals approximately 156.21 GB. The official model card lists a roughly 78 GB FP8 memory footprint and a minimum deployment configuration of two 80 GB A100s, two H100 SXM5 GPUs, one H200, one B200, or one B300. Both checkpoints are served through the Apache-2.0-licensed aleph-alpha-inference vLLM plugin, which provides Kolibri-specific reasoning and tool-call parsers.",
      "paper_url": "https://aleph-alpha.com/downloads/tech-report.pdf",
      "links": [
        "https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/",
        "https://huggingface.co/Aleph-Alpha/Kolibri-1",
        "https://huggingface.co/Aleph-Alpha/Kolibri-1-BF16",
        "https://github.com/Aleph-Alpha/aleph-alpha-inference"
      ],
      "page_url": "https://nlp.com.ai/models/kolibri-1/",
      "record_updated_at": "2026-10-11T08:37:45.844290+00:00"
    },
    {
      "slug": "naive-n05-flash",
      "source_file": "2026-09-27-naive-n0.5-flash.json",
      "release_date": "2026-09-27",
      "date": "September 27, 2026",
      "name": "Naive-N0.5-Flash",
      "developer": "NaiveAI",
      "parameters": "309B total / 15.5B active (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Open-weight coding and AI-R&D model produced by continuing from Xiaomi's MiMo-V2.5 base and replacing its global-attention layers with DeepSeek Sparse Attention. The 48-layer, 4,096-dimensional MoE contains 256 routed experts with eight selected per token and uses a hybrid stack of 39 Sliding-Window Attention layers and nine DeepSeek Sparse Attention layers, organized primarily as eight six-layer modules with five SWA layers followed by one DSA layer and an additional DSA layer at the start of the network. No layer uses full attention. SWA operates over a 128-token window, while DSA uses a 16-head lightweight indexer to select the top 2,048 historical tokens for backbone attention. DSA replaces the original MLA formulation with grouped-query attention using four KV groups; both attention mechanisms incorporate sink bias. NaiveAI reports that the redesigned indexer reduces selection latency by 44% relative to the original DSA implementation while retaining agent-task performance at the native 1,048,576-token context length. Continued training totaled 3.25 trillion tokens: 50B tokens of indexer warmup, 3T tokens of sparse-attention training, and 200B tokens of learning-rate decay. Official evaluations used Claude Code 2.1.207 with basic file and Bash tools, the full 1M-token window, temperature 1.0, and top-p 0.95. NaiveAI reports scores of 67.8 on DeepSWE v1.1, 86.7 on Terminal-Bench 2.1, 73.6 on SWE-bench Pro, 71.9 on NL2Repo-Bench, 78.2 on FrontierSWE v1, 32.4 on Agents' Last Exam CLI, 37.5 on PostTrainBench, 73.7% on MLE-bench-30, and 63.2 on PaperBench. NaiveRT, the associated deterministic inference runtime, combines mega-kernel fusion, Programmatic Dependent Launch, GPU-driven scheduling, communication overlap, and speculative decoding with a five-layer DFlash draft model. NaiveAI reports approximately 50 tokens per second per user in Standard mode and a peak single-stream decode rate of 2,122 tokens per second on eight GPUs in Ultrafast mode; the latter is the best one-second window across 41 HTML/SVG requests with thinking disabled and prefill excluded. On the same system, a speculative draft-verify-sample-commit round took 3.4 ms versus 12.3 ms with SGLang. The release provides a 49-shard BF16 checkpoint totaling approximately 617.8 GB, a 49-shard FP8 checkpoint totaling approximately 315.1 GB, and a 1.3 GB matching speculative-draft checkpoint. The repositories include custom Transformers implementations and a chat template. Although the launch describes NaiveRT as open source, the technical article says the optimized runtime repository contents and reproducibility materials are scheduled for October 12; the model and draft weights are already live. Planned API pricing is $0.10 per million input tokens, $0.40 per million output tokens, and $0.01 per million cache-read tokens.",
      "paper_url": "",
      "links": [
        "https://naive.ai/en/research/",
        "https://huggingface.co/NaiveAI/Naive-N0.5-Flash",
        "https://huggingface.co/NaiveAI/Naive-N0.5-Flash-FP8",
        "https://huggingface.co/NaiveAI/Naive-N0.5-Flash-FP8-Draft",
        "https://github.com/NaiveAI-Labs/Naive-N0.5-Flash",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Base"
      ],
      "page_url": "https://nlp.com.ai/models/naive-n05-flash/",
      "record_updated_at": "2026-10-11T08:37:45.842005+00:00"
    },
    {
      "slug": "krake-v2-legacy",
      "source_file": "2026-09-24-krake-v2-legacy.json",
      "release_date": "2026-09-24",
      "date": "September 24, 2026",
      "name": "Krake V2 Legacy",
      "developer": "NovelAI / Anlatan (United States)",
      "parameters": "21B (dense; GPT-NeoX-20B fine-tune)",
      "context_window": "2,048 tokens",
      "license": "GPL-3.0",
      "key_features": "Open-weight preservation release of NovelAI's legacy Krake V2 storytelling model ahead of its eventual retirement from the NovelAI service. Krake is a fine-tune of EleutherAI's GPT-NeoX-20B and was originally introduced on March 11, 2022, with its V2 update following on April 29, 2022. The released configuration uses a 44-layer dense GPT-NeoX causal-language-model architecture with a 6,144-dimensional hidden state, 64 attention heads, a 24,576-dimensional feed-forward layer, parallel residual connections, partial rotary embeddings, and a 2,048-token context window. NovelAI describes Krake as a temperamental but versatile storytelling model that was historically restricted to its Opus subscription tier because of its size and serving cost. For this preservation release, NovelAI converted the model from its internal format into nine FP16 Hugging Face Safetensors shards totaling approximately 41.1 GB and converted its legacy customization modules into Hugging Face PEFT format. The standard GPTNeoXForCausalLM implementation permits use through Transformers without a proprietary runtime. NovelAI says it released the checkpoint and modules for nostalgia, posterity, and historic preservation so that Krake remains available after leaving the hosted service.",
      "paper_url": "",
      "links": [
        "https://journal.novelai.net/goodbye-krake-and-clio/",
        "https://huggingface.co/NovelAI/krake-v2-legacy",
        "https://docs.novelai.net/en/text/models/"
      ],
      "page_url": "https://nlp.com.ai/models/krake-v2-legacy/",
      "record_updated_at": "2026-10-11T08:37:45.839741+00:00"
    },
    {
      "slug": "clio-v1-legacy",
      "source_file": "2026-09-24-clio-v1-legacy.json",
      "release_date": "2026-09-24",
      "date": "September 24, 2026",
      "name": "Clio V1 Legacy",
      "developer": "NovelAI / Anlatan (United States)",
      "parameters": "3B (dense)",
      "context_window": "8,192 tokens",
      "license": "GPL-2.0-only",
      "key_features": "Open-weight preservation release of Clio, NovelAI's first fully original language model pretrained from scratch, ahead of its eventual retirement from the NovelAI service. Clio was originally released on May 23, 2023 after training on NovelAI's Shoggy H100 cluster. NovelAI reports approximately 1.6 trillion training tokens, representing roughly 6 TB of text, drawn from its custom Nerdstash dataset and encoded with Nerdstash Tokenizer V1. The released checkpoint uses a 28-layer dense causal-language-model architecture compatible with Hugging Face's StableLmForCausalLM class, with a 2,816-dimensional hidden state, 22 attention and key/value heads, a 7,552-dimensional feed-forward layer, parallel residual connections, partial rotary embeddings, and an 8,192-token context window. NovelAI reports that Clio reached approximately 73% LAMBADA accuracy during pretraining and exceeded 74% during fine-tuning, while its writing performance surpassed the substantially larger Euterpe and Krake models. The Hugging Face release occupies approximately 6.36 GB across two primarily BF16 Safetensors shards, includes the original tokenizer assets, and converts NovelAI's customization modules into PEFT format. NovelAI found an existing Transformers model class that fits Clio once an additional configuration flag is enabled, so the preservation release requires no proprietary model implementation. The checkpoint and modules were published for nostalgia, posterity, and historic preservation.",
      "paper_url": "",
      "links": [
        "https://journal.novelai.net/goodbye-krake-and-clio/",
        "https://huggingface.co/NovelAI/clio-v1-legacy",
        "https://blognew.novelai.net/a-new-model-clio-is-coming-to-opus-ef4e2457c601/",
        "https://docs.novelai.net/en/text/models/"
      ],
      "page_url": "https://nlp.com.ai/models/clio-v1-legacy/",
      "record_updated_at": "2026-10-11T08:37:45.837623+00:00"
    },
    {
      "slug": "mimo-v26-mimo-v26-pro-rl-mimo-v26-flash-rl-mimo-v26-distill-qwen-9b",
      "source_file": "2026-09-22-mimo-v2.6-series.json",
      "release_date": "2026-09-22",
      "date": "September 22, 2026",
      "name": "MiMo-V2.6 (MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, MiMo-V2.6-Distill-Qwen-9B)",
      "developer": "Xiaomi MiMo (China)",
      "parameters": "MiMo-V2.6-Pro-RL: 1.02T total / 42B active (Sparse MoE); MiMo-V2.6-Flash-RL: 309B total / 15B active (Sparse MoE); MiMo-V2.6-Distill-Qwen-9B: 9B",
      "context_window": "1M tokens (Pro-RL and Flash-RL; Distill-Qwen-9B not separately specified)",
      "license": "MIT",
      "key_features": "Official open-weight MiMo-V2.6 family release led by two native omnimodal reinforcement-learning checkpoints, MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL, plus the compact MiMo-V2.6-Distill-Qwen-9B agent model. Pro-RL and Flash-RL jointly process text, images, video, and audio and support a 1M-token context window. Pro-RL uses a 1.02T-parameter sparse MoE backbone with 42B active parameters, while Flash-RL uses 309B total parameters with 15B active; both incorporate Xiaomi's MiMo ViT, separate audio tokenization and patch encoders, hybrid sliding-window and global attention, and a five-layer Multi-Token Prediction speculative decoder. Xiaomi describes the training recipe as 'You Only RL Once': one mixed reinforcement-learning run spanning coding, general agent tasks, visual interaction, and cybersecurity, with tasks and multiple agent harnesses combined within the same batches. Its fully asynchronous GRPO system processed 1,568 prompts with 16 rollouts each and approximately 3.5–3.7B tokens per training step. The release introduces Groupwise Reward Synthesis, Groupwise Advantage Redistribution, and Multi-Prefix Multi-Teacher On-Policy Distillation, while freezing the MoE router and applying adversarial reward-hacking defenses. Across six days, Flash and Pro each completed 30 RL steps, producing approximately 750,000 combined trajectories at reported training costs of $850,000 and $2.62 million respectively. Xiaomi reports that live RL raised DeepSWE v1.1 from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro; final model-card results include 67.9 and 71.9 on DeepSWE, 87.6 and 89.9 on Terminal-Bench 2.1, and 80.8 and 82.0 on OSWorld. Xiaomi additionally reports an Artificial Analysis Intelligence Index score of 46 for Pro and describes it as the strongest open-weight model at release. MiMo-V2.6-Distill-Qwen-9B is a supervised fine-tune of Qwen3.5-9B on 77.4B MiMo-generated tokens, including 27.2B loss-bearing tokens, targeting coding, general computer-use agents, visual coding, and cybersecurity.",
      "paper_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/resolve/main/MiMo_V2_6_technical_report.pdf",
      "links": [
        "https://mimo.mi.com/docs/en-US/news/latest/v2-6",
        "https://mimo.mi.com/docs/en-US/updates/model",
        "https://huggingface.co/collections/XiaomiMiMo/mimo-v26",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B",
        "https://www.modelscope.cn/models/XiaomiMiMo/MiMo-V2.6-Pro-RL",
        "https://www.modelscope.cn/models/XiaomiMiMo/MiMo-V2.6-Flash-RL"
      ],
      "page_url": "https://nlp.com.ai/models/mimo-v26-mimo-v26-pro-rl-mimo-v26-flash-rl-mimo-v26-distill-qwen-9b/",
      "record_updated_at": "2026-10-11T08:37:45.835687+00:00"
    },
    {
      "slug": "tinfield-1",
      "source_file": "2026-09-21-tinfield-1.json",
      "release_date": "2026-09-21",
      "date": "September 21, 2026",
      "name": "Tinfield 1",
      "developer": "Bad Theory Labs (Nigeria)",
      "parameters": "177B total / 6.6B active per token (Sparse MoE; approximately 180B stored parameters)",
      "context_window": "256K",
      "license": "Qwen Community License 1.0",
      "key_features": "Open-weight agentic terminal and long-horizon software-engineering model built on Qwen3.8-Flash-Next. Tinfield 1 retains the base model's sparse multimodal architecture while specializing its behavior for terminal work, repository-scale coding, and extended software-engineering trajectories. It activates approximately 6.6B parameters per token from 177B total parameters; Hugging Face metadata counts approximately 180B stored parameters across the complete checkpoint. The release provides roughly 360 GB of BF16 safetensors and a 256K context window. Bad Theory Labs evaluated Tinfield 1 with mini-swe-agent at k=5 on the complete benchmark task sets, reporting 33.0 on the 66-task Terminal-Bench 4.0 and 62.0 on the 113-task DeepSWE v1.1, compared with 29.0 and 58.7 respectively for the underlying Qwen3.8-Flash-Next model. The developer describes Tinfield 1 as the second-highest-scoring open-weight model on the contemporary Terminal-Bench leaderboard and reports results above several frontier closed models, although those leaderboard comparisons use differing model harnesses. Official local-inference derivatives include the 72 GB Tinfield 1 Compact, using IQ2_XXS gate and up projections with IQ4_NL down projections, and the 61 GB Tinfield 1 Mini, using IQ2_XXS gate and up projections with range-searched Q2_0 down projections. Bad Theory Labs states that both compressed builds can run on a single 64 GB machine, but notes that their published benchmark scores were measured using the BF16 checkpoint and that the quantized versions have not yet been evaluated on those benchmarks. An official Q8_0 GGUF build is also available.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/badtheorylabs/Tinfield-1",
        "https://huggingface.co/badtheorylabs/Tinfield-1-Compact-GGUF",
        "https://huggingface.co/badtheorylabs/Tinfield-1-Mini-GGUF",
        "https://huggingface.co/badtheorylabs/Tinfield-1-Q8_0-GGUF",
        "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
        "https://www.badtheorylabs.com/"
      ],
      "page_url": "https://nlp.com.ai/models/tinfield-1/",
      "record_updated_at": "2026-10-11T08:37:45.833767+00:00"
    },
    {
      "slug": "hemmingway-1",
      "source_file": "2026-09-20-hemmingway-1.json",
      "release_date": "2026-09-20",
      "date": "September 20, 2026",
      "name": "Hemmingway-1",
      "developer": "Altworld (Switzerland / South Africa)",
      "parameters": "27B dense (26,895,998,464 BF16 parameters)",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight, English-first writing and chat model post-trained from Qwen3.8-27B, specialized for everyday messages, emails, difficult interpersonal requests, dialogue, and creative writing that avoids sounding machine-generated or surrounding the requested text with commentary. Hemmingway-1 retains the dense 64-layer Qwen3.8-27B hybrid architecture, alternating three linear-attention layers with one full-attention layer for 48 linear and 16 global-attention layers. The official configuration uses a 5,120-dimensional hidden state, 17,408-dimensional feed-forward layers, 24 query heads, four key/value heads, a 248,320-token vocabulary, interleaved mRoPE with a 10,000,000 base, and a 262,144-token context window; a separate one-layer multi-token-prediction module is shipped alongside the twelve BF16 model shards. The included chat template supports thinking and non-thinking generation, with xhigh, medium, and low reasoning-effort settings, and includes structured tool-call formatting inherited from the Qwen family. Altworld reports that Hemmingway-1 won its internally developed CommunicationBench, consisting of eighty real writing requests evaluated through blind pairwise comparisons run in both orders, beating Fable 5.1 and finishing fifty points ahead of GPT-6 Astra. On the lab's Human-Likeness evaluation it reportedly finished twenty-six points ahead of the next model, while its category breakdown reports a 72% human-likeness preference rate on difficult asks versus 9% for GPT-6 Astra. Altworld also reports that the model commonly returns the requested message directly rather than wrapping it in options, preambles, or explanatory notes. On the externally developed EQ-Bench 4, the model card reports a third-place finish ahead of GPT-5.5, Opus 4.7, and Opus 4.8 and within twelve points of the leading model. On Altworld's internal StoryBench it reportedly performs level with Kimi K3 and 504 points above the Qwen3.8-27B starting checkpoint. The developer explicitly identifies hostile storytelling and long story turns as relative weaknesses and notes that the model is English-first and may sound confident when incorrect. The release provides standard Transformers loading instructions and an example vLLM server command using the full 262,144-token context.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Altworld/Hemmingway-1",
        "https://github.com/lukeckprobierts/Hemmingway-1",
        "https://hemmingway.io"
      ],
      "page_url": "https://nlp.com.ai/models/hemmingway-1/",
      "record_updated_at": "2026-10-11T08:37:45.831506+00:00"
    },
    {
      "slug": "aliceai-foundation-80b-a3b-base",
      "source_file": "2026-09-18-aliceai-foundation-80b-a3b-base.json",
      "release_date": "2026-09-18",
      "date": "September 18, 2026",
      "name": "AliceAI-Foundation-80B-A3B-Base",
      "developer": "Yandex (Russia)",
      "parameters": "80B total / 3B active (MoE; 81,286,433,408 stored BF16 parameters)",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight, text-only pretrained foundation model developed and trained entirely from scratch by Yandex. AliceAI-Foundation-80B-A3B-Base uses a highly sparse 48-layer hybrid causal architecture arranged as 12 repetitions of three KDA linear-attention layers followed by one gated full-attention layer, with a Mixture-of-Experts feed-forward block attached to every layer. The model has a 2,048-dimensional hidden state, a 129,024-token vocabulary, and approximately 80B marketed parameters while activating about 3B per token; the published safetensors metadata counts exactly 81,286,433,408 BF16 parameters. Each MoE layer contains 512 routed experts, selects ten experts per token, and additionally activates one shared expert, with 512-dimensional expert intermediate states, sigmoid routing, and router-bias correction. Its 36 KDA layers use 32 query, key, and value heads with 128-dimensional heads and a four-token convolution kernel, while the 12 gated-attention layers use 16 query heads, two KV heads, 256-dimensional heads, and partial rotary position encoding. The architecture also contains a single Multi-Token Prediction layer for speculative decoding and supports a native 262,144-token context window. Yandex assembled a new Russian-and-English training corpus containing data for factual knowledge, complex reasoning, coding, and tool use. The final pretraining-token count is not disclosed, but Yandex says key architectural and hyperparameter decisions were validated through multiple independent training runs from scratch, each consuming two trillion tokens. This release represents the pretraining stage and deliberately does not define a default chat template; raw generation should not be treated as instruction-following behavior. The repository instead includes the agentic conversation format encountered during training, a separate Jinja template for preparing supervised fine-tuning data, and a four-80GB-GPU FSDP2 LoRA example. Vendor evaluations were conducted through Yandex's internal infrastructure using vLLM. At deterministic temperature zero, Yandex reports 86.5 on WikiWebFacts, 67.9 on HardMultiQA, 86.5 on CultCat, 74.2 on EduBench Russian, 90.5 on the Russian Unified State Exam benchmark, 66.8 on MMLU-Pro, 44.3 on SuperGPQA, 48.3 on BigCodeBench, 50.5 on LiveCodeBench v5-6 with one-shot chain of thought, 74.1 on FinQA 128K, and 64.6 on LongMemEval 128K. In higher-temperature pass-at-k evaluations, it reports 96.7 pass@32 on AIME 2026, 96.9 pass@32 on HMMT February 2026, 88.7 pass@8 on IMO AnswerBench, 60.4 pass@1 and 82.9 pass@8 on LiveCodeBench v5-6, and 68.9 pass@8 on Codeforces C++. The release is particularly strong on Yandex's Russian factual and educational evaluations, while its English TriviaQA and several general-knowledge results remain behind larger-active-parameter baselines. Yandex simultaneously released the WikiWebFacts and HardMultiQA Russian factual-knowledge datasets and their evaluation protocols. The BF16 checkpoint is distributed across 49 safetensors shards totaling approximately 162.58 GB. Transformers loading uses bundled custom AliceAI model code and therefore requires trust_remote_code=True, while the reference vLLM launch enables the included one-token MTP speculative decoder.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base",
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base/raw/main/README_en.md",
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base/raw/main/config.json",
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base/raw/main/LICENSE",
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base/tree/main/finetune",
        "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base/commits/main",
        "https://huggingface.co/datasets/yandex/WikiWebFacts",
        "https://huggingface.co/datasets/yandex/HardMultiQA"
      ],
      "page_url": "https://nlp.com.ai/models/aliceai-foundation-80b-a3b-base/",
      "record_updated_at": "2026-10-11T08:37:45.829621+00:00"
    },
    {
      "slug": "xing40-29b-a4b",
      "source_file": "2026-09-17-xing4.0-29b-a4b.json",
      "release_date": "2026-09-17",
      "date": "September 17, 2026",
      "name": "Xing4.0-29B-A4B",
      "developer": "China Telecom Artificial Intelligence Technology Co., Ltd. (China)",
      "parameters": "29B total / 4B active (MoE)",
      "context_window": "262,144 tokens native; extensible to 524,288 tokens",
      "license": "Apache-2.0",
      "key_features": "Official open-weight, text-only, agent-oriented language model and the latest generation of the Xing series, formerly known as TeleChat. Xing4.0-29B-A4B uses a custom 40-layer Mixture-of-Experts architecture combining manifold-constrained Hyper-Connections (mHC), Multi-head Latent Attention (MLA), and Multi-Token Prediction (MTP). Its shipped configuration specifies a 3,584-dimensional hidden state, two initial dense layers, 64 routed experts plus one shared expert, four routed experts selected per token, 1,024-dimensional expert feed-forward layers, a 9,216-dimensional dense feed-forward size, mHC multiplicity four, and one next-token prediction layer. The model has 29B total parameters while activating approximately 4B per token. It natively supports 262,144-token contexts and is advertised as extensible to 524,288 tokens using YaRN-based position scaling. The model supports multi-step planning, structured tool calling, long-context reasoning, and switchable thinking through the enable_thinking chat-template option. China Telecom reports that Xing4.0 was trained entirely on Ascend 910C NPUs using MindSpore and MindFormers, making it the first model at this scale trained entirely on the domestic Ascend software and hardware stack. Fine-grained MoE communication optimization, selective recomputation, automatic graph and operator fusion, and custom fused mHC operators reportedly improved training throughput by approximately 96% over the unoptimized implementation. Vendor-reported scores include 75.0 on SWE-bench Verified, 66.0 on SWE-bench Multilingual, 57.5 on Terminal-Bench 2.1, 76.55 on Claw-Eval, 64.63 on Tau3-Bench, 60.8 on DeepResearchBII, 90.0 on AIME 2026, 69.67 on IFBench, and 61.0 on AA-LCR. The evaluations use long-running agent harnesses and unusually large inference budgets, including a 24-hour Terminal-Bench timeout, 210K-token SWE-agent contexts, and up to 131,072 output tokens on AIME. The BF16 repository totals approximately 62.4 GB across 41 safetensors shards, while official FP8 and GGUF variants were released concurrently. Transformers loading currently requires trust_remote_code=True because the checkpoint ships custom Xing4.0 configuration and modeling code. Official deployment support includes vLLM, SGLang, and KTransformers; fine-tuning support includes LLaMA-Factory and MindFormers; and the release reports targeted format alignment for OpenCode, Claude Code, OpenClaw, and Hermes.",
      "paper_url": "",
      "links": [
        "https://github.com/XingChen-AGI/Xing4.0-29B-A4B",
        "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B",
        "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B/blob/main/config.json",
        "https://modelscope.cn/models/XingChen-AGI/Xing4.0-29B-A4B",
        "https://modelers.cn/models/XingChen-AGI/Xing4.0-29B-A4B",
        "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-FP8",
        "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF",
        "https://github.com/sgl-project/sglang/pull/39793"
      ],
      "page_url": "https://nlp.com.ai/models/xing40-29b-a4b/",
      "record_updated_at": "2026-10-11T08:37:45.827949+00:00"
    },
    {
      "slug": "ternary-bonsai-2-27b",
      "source_file": "2026-09-17-bonsai-2-27b.json",
      "release_date": "2026-09-17",
      "date": "September 17, 2026",
      "name": "Ternary Bonsai 2 27B",
      "developer": "Prism ML, Inc. (United States)",
      "parameters": "27.36B total: 24.35B language backbone + 2.54B embeddings/LM head + 0.46B vision tower",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Second-generation open-weight, aggressively compressed vision-language model from Prism ML, derived from Qwen3.8-27B with the base architecture unchanged. Ternary Bonsai 2 27B retains the base model's 64-block hybrid-attention language backbone, comprising approximately 75% linear-attention and 25% full-attention layers with SwiGLU feed-forward networks, RoPE, RMSNorm, native reasoning, text-and-image input, structured tool calling, and a 262,144-token context window. Prism ML's detailed model card counts 27.36B parameters, comprising a 24.35B language backbone, 2.54B embedding and language-model-head parameters, and an optional 0.46B, 27-block vision tower; the launch press release rounds the model to 27.8B. Language-model weights are represented end to end as ternary {-1, 0, +1} values with one shared FP16 scale per group of 128 weights, covering embeddings, attention projections, MLP projections, and the LM head. A fixed blockwise Hadamard rotation with block size 1,024 is folded into the stored weights, while the corresponding transform is applied to activations at runtime. Only 26.2M language-model parameters, approximately 0.098%, remain in higher precision. The compact PTQ1_0 GGUF densely packs ternary values at 1.75 effective bits per weight into a 5.95 GB language-model file, while the PQ2_0 version uses two-bit slots and occupies 7.21 GB in exchange for less expensive unpacking and substantially faster prompt processing on some hardware. Optional vision projectors add approximately 629 MB in Q8_0 or 931 MB in BF16. An MLX companion package for Apple Silicon occupies approximately 8.6 GB including the language model and vision tower. Prism ML reports an overall score of 83.9 across a 20-benchmark thinking-mode suite, retaining 98.2% of the full-precision Qwen3.8-27B baseline's 85.4. Reported category scores include 96.57 for math, 81.58 for coding, 82.66 for instruction following, 77.57 for agentic and tool-calling tasks, 83.95 for knowledge and reasoning, and 78.59 for vision. A narrower 14-benchmark table in the current model card reports an 84.78 average versus 86.32 for FP16, including 95.83 on AIME 2026, 90.07 on LiveCodeBench, 74.92 on BFCL v3, 75.49 on MMMU-Pro, and 74.00 on IFBench. Prism ML reports throughput of up to 143 tokens per second on an RTX 5090 and 46.8 tokens per second on an Apple M5 Max, with custom CUDA and Metal low-bit kernels. The packed GGUF formats require Prism ML's modified llama.cpp runtime because stock llama.cpp does not yet implement the required ternary hybrid-attention kernels and rotated-weight metadata; the MLX release likewise uses Prism ML's custom MLX stack. Official integrations and examples cover OpenAI-compatible serving, Pi, Hermes Agent, OpenClaw, Cline, multimodal workflows, MCP tools, and per-request reasoning budgets.",
      "paper_url": "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-27b-whitepaper.pdf",
      "links": [
        "https://prismml.com/news/bonsai-2-27b",
        "https://huggingface.co/collections/prism-ml/bonsai-2",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit",
        "https://github.com/PrismML-Eng/Bonsai-demo",
        "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-27b-whitepaper.pdf",
        "https://github.com/PrismML-Eng/llama.cpp",
        "https://github.com/PrismML-Eng/mlx",
        "https://huggingface.co/Qwen/Qwen3.8-27B"
      ],
      "page_url": "https://nlp.com.ai/models/ternary-bonsai-2-27b/",
      "record_updated_at": "2026-10-11T08:37:45.826101+00:00"
    },
    {
      "slug": "zero-mc001",
      "source_file": "2026-09-15-zero-mc001.json",
      "release_date": "2026-09-15",
      "date": "September 15, 2026",
      "name": "Zero (MC001)",
      "developer": "Moving Castles",
      "parameters": "8B (dense)",
      "context_window": "32,768 tokens native; production deployment uses 16,384",
      "license": "Not specified",
      "key_features": "Open-weight experimental character model created by fine-tuning Qwen3-8B-Base to embody one narrow, persistent, anti-servile character rather than a general-purpose assistant. Zero represents a man voluntarily enclosed in a white plastic box with no personal memories before November 5, 2021 and only five objects in his environment. Moving Castles defines the character through a human-authored association graph containing 117 nodes and 298 edges, supplemented by separate core-personality and writing-style anchors. Its synthetic supervised corpus was generated through multi-turn conversations between independently sampled model families, without system messages, and contains 5,932 post-filter conversations, 73,765 character turns, and approximately 3.04 million character tokens. Conversations span 10 to 40 turns so that long-horizon character persistence remains in distribution. Supervised training began from the non-instruction-tuned Qwen3-8B-Base checkpoint and used three epochs of rank-64, alpha-128 LoRA training together with fully trained token embeddings and LM head, for 1.42B trainable parameters; the adapter was subsequently merged. Reinforcement learning then used 380 prompts harvested from observed character failures and successes, comprising 193 capitulation onsets, 107 pressure-held examples, and 80 ordinary in-world openers. This stage ran 300 steps of GRPO with DAPO loss modifications, four rollouts per prompt, a rank-16, alpha-32 LoRA adapter, and 43.6M trainable parameters. Its reward combines a Qwen3.6-27B LLM judge scoring voice, world frame, personality, and engagement against the character bible with a penalty for excessive four-gram repetition. In a held-out evaluation of 250 matched conversations with 16 turns each, the merged SFT-plus-RL model exhibited hard character breaks in 2.8 percent of conversations, compared with 22.8 percent for the SFT-only checkpoint and 45.2 percent for the sibling Qwen3-8B instruct model controlled through a system prompt. The final model had zero complete unrecovered collapses, compared with 21 for SFT-only and 60 for prompting. This specialization substantially damages conventional capabilities: Zero scored 2.5 percent on GSM8K versus 62.2 percent for Qwen3-8B-Base, and 20.7 percent prompt-level and 32.7 percent instruction-level strict IFEval versus 46.8 and 58.2 percent for the base model. Mechanistic experiments report that the residual-stream displacement produced by SFT and RL is highly aligned at mean cosine 0.985, while the trained character directions are nearly orthogonal to the system-prompt direction at -0.022 and -0.004. Steering the base model with the fine-tuning displacement elicited traces of Zero, whereas steering with the prompt displacement did not, although the authors caution that orthogonality alone is weak evidence in high-dimensional space. The model should be used without a system prompt because training includes only user and assistant turns; it also requires both ChatML end-of-turn and end-of-text tokens as stopping conditions at nonzero temperature. The checkpoint is distributed in BF16 across four safetensors shards and uses the standard Qwen3 architecture. No explicit license is currently declared for the released checkpoint.",
      "paper_url": "https://movingcastles.world/posts/zero",
      "links": [
        "https://huggingface.co/movingcastles/zero",
        "https://movingcastles.world/posts/zero",
        "https://huggingface.co/movingcastles/zero/blob/main/config.json",
        "https://huggingface.co/Qwen/Qwen3-8B-Base"
      ],
      "page_url": "https://nlp.com.ai/models/zero-mc001/",
      "record_updated_at": "2026-10-11T08:37:45.823847+00:00"
    },
    {
      "slug": "atria-dawn-preview-bf16-fp8",
      "source_file": "2026-09-14-atria-dawn-preview.json",
      "release_date": "2026-09-14",
      "date": "September 14, 2026",
      "name": "Atria Dawn Preview (BF16, FP8)",
      "developer": "Shanghai Artificial Intelligence Laboratory (China)",
      "parameters": "753B checkpoint / ~40B active (MoE; post-trained from the 744B-A40B GLM-5.2 foundation)",
      "context_window": "256K supported release context (checkpoint max_position_embeddings: 1,048,576)",
      "license": "MIT",
      "key_features": "Open-weight preview agentic model developed by the Shanghai Artificial Intelligence Laboratory for long-horizon scientific research, software engineering, digital work, and tool-mediated problem solving. Atria Dawn Preview is a post-trained derivative of Z.AI's GLM-5.2 foundation rather than a newly pretrained base model. Its shipped configuration retains the 78-layer GLM MoE DSA architecture, comprising three initial dense layers followed by 75 sparse layers, a 6,144-dimensional hidden state, 256 routed experts plus one shared expert, eight routed experts selected per token, 64 attention heads, compressed query and key/value representations, a 2,048-token sparse-attention index, shared indexers across groups of sparse layers, and one Multi-Token Prediction layer. Hugging Face calculates approximately 753B checkpoint parameters, while the developers describe the underlying GLM-5.2 foundation as 744B parameters; the inherited architecture activates approximately 40B parameters per token. The model was trained through a Verifiable Experience Pipeline in which agents interact with real execution environments, call tools, produce intermediate artifacts, respond to environmental feedback, and have final outcomes checked through executable tests, experiment metrics, file and application state, geometric checks, source evidence, or human-defined criteria. Trajectory curation removes incomplete, contradictory, duplicate, and behaviorally invalid examples, while failure analysis informs additional task construction and environment refinement. The developers organize its target capabilities into discovery, creation, delivery, and authorized cybersecurity work. They report leading scores within their comparison table on AutomationBench at 53.8, BFCL v4 at 77.0, CyberGym at 86.5, DeepSearchQA at 96.0, and BrowseComp at 92.5. Additional reported results include 66.4 on SkillsBench, 65.0 on Workspace-Bench, 86.2 HumanRank on MLE-bench Lite, 78.3 on Terminal-Bench 2.1, 59.6 on SWE-bench Pro, 51.1 on DeepResearch Bench II, and 1583 Elo on GDPval. The accompanying report analyzes 769 development-task records from 56 participants; among recorded methods and decisions, agents proposed 64.6 percent while humans retained the final choice in 85.5 percent. The public release and evaluation harnesses use a 256K-token context, although the checkpoint configuration declares 1,048,576 maximum positions. Official BF16 and FP8 checkpoints occupy approximately 1.51 TB and 756 GB respectively, with deployment support documented for SGLang 0.5.13.post1 or newer and vLLM 0.23.0 or newer. The model accepts text input and produces text output.",
      "paper_url": "https://github.com/atria-asi/Atria-Dawn-Preview/blob/main/atria-dawn-preview.pdf",
      "links": [
        "https://atria-asi.ai/",
        "https://github.com/atria-asi/Atria-Dawn-Preview",
        "https://huggingface.co/internlm/Atria-Dawn-Preview",
        "https://huggingface.co/internlm/Atria-Dawn-Preview-FP8",
        "https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Atria-Dawn-Preview",
        "https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Atria-Dawn-Preview-FP8",
        "https://huggingface.co/internlm/Atria-Dawn-Preview/blob/main/config.json",
        "https://github.com/atria-asi/Atria-Dawn-Preview/blob/main/atria-dawn-preview.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/atria-dawn-preview-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.821653+00:00"
    },
    {
      "slug": "agnes-30-flash",
      "source_file": "2026-09-11-agnes-3.0-flash.json",
      "release_date": "2026-09-11",
      "date": "September 11, 2026",
      "name": "Agnes-3.0-Flash",
      "developer": "Agnes AI (Singapore)",
      "parameters": "33.09B (dense)",
      "context_window": "262,144 tokens",
      "license": "Apache License 2.0",
      "key_features": "Open-weight native multimodal reasoning model designed for coding, instruction following, long-context work, and tool-driven agents. Agnes-3.0-Flash accepts text, images, and video and generates text, with a 33.09-billion-parameter dense decoder and an approximately 66 GB BF16 checkpoint. Its hybrid architecture contains 72 decoder layers arranged in a repeating 3:1 pattern: 54 gated delta-rule recurrent layers and 18 global-attention layers, meaning only one quarter of the layers maintain a KV cache that grows with context. The model uses a 5,120-dimensional hidden state, 24 query heads and four KV heads in global-attention layers, causal-convolutional gated delta layers, SwiGLU feed-forward networks with an additional parallel 2,048-wide SwiGLU branch, and a 248,320-token vocabulary. Its vision tower has 27 layers, a 1,152-dimensional hidden state, 16-by-16 patches, 2-by-2 spatial merging, and projection into the language model's hidden space. The supplied chat template supports xhigh, medium, and low reasoning effort plus thinking-disabled operation, structured tool calling, and multimodal conversation. The checkpoint requires its bundled custom Transformers implementation with trust_remote_code enabled; Agnes also supplies patches and a launch script for SGLang. Agnes reports 74.20 on IFBench, 38.08 on SciCode, 85.05 on GPQA Diamond, 68.33 on AA-LCR, and 23.00 on AA Omniscience Accuracy, while cautioning that its comparison table combines different sources, harnesses, model snapshots, and parameter conventions rather than a controlled head-to-head evaluation. The official BF16 deployment guidance recommends one H100 80GB or H200 141GB and at least 128GB of host memory.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Agnes-AI/Agnes-3.0-Flash",
        "https://huggingface.co/Agnes-AI/Agnes-3.0-Flash/raw/main/README.md",
        "https://huggingface.co/Agnes-AI/Agnes-3.0-Flash/blob/main/config.json",
        "https://huggingface.co/Agnes-AI/Agnes-3.0-Flash/blob/main/LICENSE",
        "https://agnes-ai.com/"
      ],
      "page_url": "https://nlp.com.ai/models/agnes-30-flash/",
      "record_updated_at": "2026-10-11T08:37:45.819386+00:00"
    },
    {
      "slug": "deepseek-v41-flash",
      "source_file": "2026-09-10-deepseek-v4.1-flash.json",
      "release_date": "2026-09-10",
      "date": "September 10, 2026",
      "name": "DeepSeek-V4.1-Flash",
      "developer": "DeepSeek AI (China)",
      "parameters": "552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Official open-weight multimodal model released immediately after a two-day first-party API beta. DeepSeek-V4.1-Flash is trained from scratch rather than being another post-training update to V4-Flash. It uses a 40-layer Causal Encoder-Decoder architecture comprising a 20-layer causal encoder and 20-layer decoder, with 552B backbone parameters and 196B additional Engram conditional-memory parameters. The model activates 8B parameters per token during prefill and 16B during decoding, using one shared expert and 384 routed experts per MoE layer with six routed experts selected per token. Compressed Sparse Attention 2 introduces Full, Reindex, and Reuse attention modes, cross-layer KV and index reuse, and hierarchical sparse indexing. Combined with FP4 main-KV caching, this reduces the always-resident global KV cache to 890 bytes per token, approximately one quarter of DeepSeek-V4-Flash, while SWA Bounded Replay reduces the persistent SSD or host-memory KV footprint to approximately one eighth. Additional components include Single-Pass mHC residual mixing, two sparsely accessed Engram modules, and integrated DSpark speculative decoding. A 32-layer DeepSeek-ViT vision encoder with 2D RoPE, 14-pixel patches, 3x3 pixel-unshuffle downsampling, and up to 1,024 image tokens provides native image understanding. Pretraining used a 45T-token multimodal corpus, with sparse attention trained from scratch at 64K sequence length and context extended to 1,048,576 tokens after 34T tokens. Post-training follows SFT, reinforcement learning, and on-policy distillation while scaling automatically synthesized agent tasks, environments, and rollouts; reasoning effort is continuously controllable from 1 to 100. At maximum effort, DeepSeek reports 90.6 on Terminal-Bench 2.1, 30.0 on Terminal-Bench 3.0, 31.2 on Terminal-Bench 4.0, 74.2 on DeepSWE v1.1 with mini-SWE-agent, 72.6 with DeepSeek Harness Minimal, 88.1 on CyberGym, 54.8 on AutomationBench, and 31.8 on Agent's Last Exam. The checkpoint is published across 48 mixed FP4/FP8 safetensors shards totaling approximately 510 GB. It does not include a Jinja chat template; DeepSeek instead provides reference conversation encoding, maintained deepseek-recipe libraries, local-inference guidance, and reproducible DeepSWE evaluation instructions.",
      "paper_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/config.json",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/tree/main/encoding",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/tree/main/inference",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/tree/main/evaluation",
        "https://github.com/deepseek-ai/deepseek-recipe"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v41-flash/",
      "record_updated_at": "2026-10-11T08:37:45.816854+00:00"
    },
    {
      "slug": "nex-n25-max-mini",
      "source_file": "2026-09-08-nex-n2.5.json",
      "release_date": "2026-09-08",
      "date": "September 8, 2026",
      "name": "Nex-N2.5 (Max, mini)",
      "developer": "Nex AGI (China)",
      "parameters": "1.6T total (Max); 35B total (mini)",
      "context_window": "1,048,576 tokens configured for Max (official serving recipe uses 262,144); 262,144 tokens for mini",
      "license": "Apache-2.0",
      "key_features": "Official open-weight portion of the Nex-N2.5 agent-model family, covering the downloadable Nex-N2.5-Max and Nex-N2.5-mini checkpoints. The same-day Nex-N2.5-Pro variant is announced and available through hosted inference, but is excluded from this release grouping because its official repositories currently contain only a model card and figures rather than checkpoint shards or a configuration. Nex-N2.5-Max is a text-only 1.6-trillion-parameter Mixture-of-Experts model based on a DeepSeek-V4-style architecture and represents Nex AGI's first complete trillion-scale post-training effort. Its configuration specifies 61 layers, a 7,168-dimensional hidden state, 384 routed experts plus one shared expert, six selected routed experts per token, three initial hash-MoE layers, Hyper Connections with multiplicity four, and alternating heavily compressed and compressed sparse attention. The checkpoint configures 1,048,576 positions using YaRN factor 16 from an original 65,536-token window, although Nex AGI's published multi-node serving recipe currently caps runtime context at 262,144. Expert weights use blockwise FP8 while other components remain at higher precision; the repository totals approximately 1.65 TB across 644 safetensors shards. Nex-N2.5-mini is a 35B multimodal Qwen3.5-style sparse MoE with 40 language layers, three linear-attention layers for every full-attention layer, 256 routed experts with eight selected per token, a shared expert, and a native 262,144-token context. Its visual stack uses a 27-layer, 1,152-dimensional ViT with spatial and temporal patching, enabling image-grounded computer use and agent workflows. The mini checkpoint totals approximately 70.2 GB across 16 BF16 safetensors shards. Both released variants support non-thinking, adaptive-thinking, and always-thinking modes through reasoning_effort, along with structured function calling for agent harnesses. Nex AGI's vendor-reported results include 86.1 on Terminal-Bench 2.1, 65.7 on SWE-Bench Pro, 65.6 on DeepSWE v1.1, and 92.6 on BrowseComp for Max; mini reports 73.4, 43.8, 36.1, and 83.4 respectively, plus 71.2 on OSWorld-Verified and 52.9 on Vision2Web. These scores use Nex AGI's stated NexAU and NexCUA harnesses and should be interpreted as vendor evaluations rather than controlled independent comparisons.",
      "paper_url": "",
      "links": [
        "https://github.com/nex-agi/Nex-N2.5",
        "https://huggingface.co/collections/nex-agi/nex-n25",
        "https://huggingface.co/nex-agi/Nex-N2.5-Max",
        "https://huggingface.co/nex-agi/Nex-N2.5-Max/blob/main/config.json",
        "https://modelscope.cn/models/nex-agi/Nex-N2.5-Max",
        "https://huggingface.co/nex-agi/Nex-N2.5-mini",
        "https://huggingface.co/nex-agi/Nex-N2.5-mini/blob/main/config.json",
        "https://modelscope.cn/models/nex-agi/Nex-N2.5-mini"
      ],
      "page_url": "https://nlp.com.ai/models/nex-n25-max-mini/",
      "record_updated_at": "2026-10-11T08:37:45.814881+00:00"
    },
    {
      "slug": "ling-30-flash-vl",
      "source_file": "2026-09-08-ling-3.0-flash-vl.json",
      "release_date": "2026-09-08",
      "date": "September 8, 2026",
      "name": "Ling-3.0-flash-VL",
      "developer": "inclusionAI (China)",
      "parameters": "124B total / 5.5B active (MoE)",
      "context_window": "Up to 1M tokens (131,072 native configuration; official 262,144-token YaRN recipe)",
      "license": "MIT",
      "key_features": "Official open-weight native multimodal extension of Ling-3.0-flash, adding image and video inputs to the model's text generation, reasoning, tool use, and agentic capabilities. inclusionAI positions vision as part of the complete understanding, planning, acting, and verification loop rather than only a perception interface, with reported support for complex layouts, charts, documents, long-video question answering, temporal event localization, visual verification, software interfaces, and visually grounded action generation. The visual stack uses a 27-layer ViT encoder with a 1,152-dimensional hidden state, 16 attention heads, 16-by-16 patches, temporal patching, and a two-layer MLP projector, while VideoRoPE represents both spatial position and temporal order. The 42-layer language backbone alternates 35 Kimi Delta Attention layers with seven Gated Multi-head Latent Attention layers in a repeating 5:1 pattern. Its sparse Mixture-of-Experts architecture contains 512 routed experts, selects eight per token, begins with two dense layers, and activates a marketed 5.5B of 124B total parameters per token. The model card advertises context up to 1M tokens, although the checkpoint configuration specifies 131,072 native positions and the published SGLang recipe demonstrates 262,144 tokens using YaRN factor 2. Thinking is enabled by default but can be disabled per request, and the supplied chat template supports structured tool calling and image or video message content. inclusionAI reports a score of 42 on Artificial Analysis Intelligence Index v4.1.1, four points above the text-only Ling-3.0-flash result reported on the same version. The release includes a roughly 250 GB BF16 checkpoint split across 64 safetensors shards and a same-day official FP8 checkpoint, with deployment guidance for SGLang and inclusionAI's Ling-compatible vLLM fork.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-VL",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-fp8",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-VL/blob/main/config.json",
        "https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash-VL",
        "https://github.com/inclusionAI/vllm-ling-v3"
      ],
      "page_url": "https://nlp.com.ai/models/ling-30-flash-vl/",
      "record_updated_at": "2026-10-11T08:37:45.812623+00:00"
    },
    {
      "slug": "minicpm5-2b-final-sft-midtrain-base-gguf-mlx-gptq-dspark",
      "source_file": "2026-09-07-minicpm5-2b.json",
      "release_date": "2026-09-07",
      "date": "September 7, 2026",
      "name": "MiniCPM5-2B (Final, SFT, Midtrain, Base, GGUF, MLX, GPTQ, DSpark)",
      "developer": "OpenBMB / ModelBest (China)",
      "parameters": "2,516,756,480",
      "context_window": "131,072 tokens",
      "license": "Apache 2.0",
      "key_features": "Same-day open-weight release of the complete MiniCPM5-2B family, comprising the fully post-trained BF16 model, SFT-only checkpoint, mid-training checkpoint, pretrained Base model, GGUF and 4-bit MLX deployment variants, a 4-bit GPTQ checkpoint, and a DSpark speculative-decoding draft model. MiniCPM5-2B is a dense causal language model using the standard LlamaForCausalLM architecture with exactly 2,516,756,480 total parameters, 1,981,982,720 non-embedding parameters, 42 layers, a 2,048-dimensional hidden state, 16 query heads, two key/value heads, grouped-query attention, and a native 131,072-token context window. The final checkpoint is designed for local assistants, coding agents, reasoning, long-context work, and tool-using workflows. Its chat template supports both thinking and no-thinking generation from the same checkpoint, as well as native XML-style function calls that SGLang's minicpm5 parser converts into OpenAI-compatible tool calls. The standard architecture requires no custom model code or kernels and is supported through Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, and multiple AI-chip backends through FlagOS. OpenBMB describes a three-stage training pipeline covering base training, mid-training, and post-training. Post-training begins with 400B tokens of deep-thinking SFT, continues with specialized reinforcement-learning teachers for mathematics, coding, agentic tasks, writing, and other domains, and concludes with On-Policy Distillation that consolidates 16 expert models, including five agentic experts, into the final checkpoint. OpenBMB simultaneously released 500K agent-training samples and an 85,995-sample verifiable RL collection spanning mathematics, code, long-context understanding, and scientific knowledge. In the developer's evaluation table, MiniCPM5-2B records a 53.9 aggregate average and particularly strong results for its size in coding, mathematics, long-context retrieval, tool use, search, and agent tasks, including 69.1 on LiveCodeBench v6, 86.5 on AIME 2026, 68.1 on NoLiMa, 66.6 on BFCL v4, 46.4 on SWE-bench Verified, 43.5 on BrowseComp-ZH, and 88.7 on GAIA Text-103. The card identifies which results came from Artificial Analysis and which were reproduced internally.",
      "paper_url": "https://arxiv.org/pdf/2506.07900",
      "links": [
        "https://huggingface.co/openbmb/MiniCPM5-2B",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B",
        "https://huggingface.co/openbmb/MiniCPM5-2B/blob/main/config.json",
        "https://huggingface.co/openbmb/MiniCPM5-2B-SFT",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-SFT",
        "https://huggingface.co/openbmb/MiniCPM5-2B-Midtrain",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-Midtrain",
        "https://huggingface.co/openbmb/MiniCPM5-2B-Base",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-Base",
        "https://huggingface.co/openbmb/MiniCPM5-2B-GGUF",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-GGUF",
        "https://huggingface.co/openbmb/MiniCPM5-2B-MLX",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-MLX",
        "https://huggingface.co/openbmb/MiniCPM5-2B-GPTQ",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-GPTQ",
        "https://huggingface.co/openbmb/MiniCPM5-2B-DSpark",
        "https://www.modelscope.cn/models/OpenBMB/MiniCPM5-2B-DSpark",
        "https://huggingface.co/datasets/openbmb/UltraData-SFT-Agent-2609",
        "https://huggingface.co/datasets/openbmb/UltraData-RL-2609",
        "https://github.com/OpenBMB/MiniCPM"
      ],
      "page_url": "https://nlp.com.ai/models/minicpm5-2b-final-sft-midtrain-base-gguf-mlx-gptq-dspark/",
      "record_updated_at": "2026-10-11T08:37:45.810570+00:00"
    },
    {
      "slug": "ling-30-flash-fin",
      "source_file": "2026-09-03-ling-3.0-flash-fin.json",
      "release_date": "2026-09-03",
      "date": "September 3, 2026",
      "name": "Ling-3.0-flash-Fin",
      "developer": "inclusionAI (China)",
      "parameters": "124B total / 5.1B active (MoE; 127.49B stored parameters)",
      "context_window": "262,144 tokens",
      "license": "MIT",
      "key_features": "The first finance-enhanced model in the Ant Ling family, developed by Ant Group with financial institutions and domain experts as a continued-training derivative of Ling-3.0-flash. Ling-3.0-flash-Fin retains the base model's 42-layer BailingMoeV3 hybrid-linear Mixture-of-Experts architecture, comprising 35 Kimi Delta Attention layers and seven gated Multi-head Latent Attention layers in a repeating 5:1 pattern. It has two initial dense layers, 512 routed experts plus one shared expert, activates eight routed experts per token, and uses a 2,560-dimensional hidden state with 32 attention heads. The model is marketed as containing 124B total and 5.1B active parameters, while its Hugging Face safetensors metadata counts 127.49B stored parameters. Its configuration supports a 262,144-token context window. inclusionAI continued training the model on high-quality financial data to support end-to-end financial research workflows that combine information retrieval, source review, calculation, modeling, and report preparation. Targeted capabilities include reconciling figures, definitions, assumptions, and reporting periods across annual reports, earnings releases, regulatory filings, and research documents; source-grounded financial search; valuation modeling; spreadsheet formulas and cross-sheet dependencies; actual-versus-estimate updates; balance checks; scenario analysis; and production of editable research materials. The accompanying FinFIRST benchmark contains 123 bilingual Chinese and English financial-search problems with detailed rubrics, source requirements, calculations, and temporal constraints. In inclusionAI's reported evaluations, Ling-3.0-flash-Fin scores 52.85 on FinFIRST, 77.04 on FinSearchComp Verified, 51.74 on FinCRAFT, 69.19 on Finance Agent v1.1, 59.81 on Finance Agent v2, 29.17 on APEX-Agents, 86.50 on SpreadsheetBench V1, 21.81 on SpreadsheetBench V2, and 41.00 on tau3-Banking. The evaluations use different benchmark-specific harnesses: financial-search tests use ReAct with web search, page visits, and Python; SpreadsheetBench uses a Claude Code harness with long turn and time limits; and the tau3-Banking result is sourced from Artificial Analysis. Thinking is enabled by default, the chat template supports disabling it and structured tool calls, and inclusionAI recommends temperature 1.0, top-p 0.95, and top-k 20 for general inference. The initial open checkpoint is released only in BF16 as 64 main weight shards plus an MTP checkpoint and is compatible with the same SGLang and Ling-specific vLLM runtimes as Ling-3.0-flash. inclusionAI cautions that financial assumptions, valuations, and investment conclusions require professional review and says future releases will explore finance-enhanced models at larger scales.",
      "paper_url": "",
      "links": [
        "https://x.com/AntLingAGI/status/2093022087069958492",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-Fin",
        "https://huggingface.co/collections/inclusionAI/ling-30-fin",
        "https://huggingface.co/datasets/inclusionAI/FinFIRST",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash",
        "https://developer.ant-ling.com/en/docs/models/ling/",
        "https://github.com/inclusionAI/vllm-ling-v3"
      ],
      "page_url": "https://nlp.com.ai/models/ling-30-flash-fin/",
      "record_updated_at": "2026-10-11T08:37:45.808237+00:00"
    },
    {
      "slug": "k2-horizon-09b-37b-7b-32b-stage1-36b-a4b-375b-a23b",
      "source_file": "2026-09-03-k2-horizon.json",
      "release_date": "2026-09-03",
      "date": "September 3, 2026",
      "name": "K2 Horizon (0.9B, 3.7B, 7B, 32B Stage1, 36B-A4B, 375B-A23B)",
      "developer": "Institute of Foundation Models / MBZUAI (United Arab Emirates)",
      "parameters": "1.08B, 5.06B, 9.00B, and 34.78B stored parameters for the dense models marketed as 0.9B, 3.7B, 7B, and 32B; 37.44B / 4B active and 379.17B / 23B active for the MoE models",
      "context_window": "131,072 tokens (0.9B); 524,288 tokens (3.7B through 375B)",
      "license": "Apache 2.0",
      "key_features": "A same-day family of six open-weight text-generation foundation models spanning edge deployment through frontier-scale agentic workloads. The family contains dense 0.9B, 3.7B, 7B, and currently released 32B Stage1 models, a 36B mixture-of-experts model with 4B active parameters and Mixture of Value Attention (MoVA), and a flagship 375B mixture-of-experts model with 23B active parameters. The 3.7B-through-375B variants support native 512K-token contexts, while the 0.9B model supports 128K with YaRN. IFM says the models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, with a smaller vocabulary for the 0.9B variant. The release introduces diffusion distillation for generating blocks of tokens in parallel, with IFM claiming roughly three-times faster generation without quality loss, plus dynamic routing across model sizes. The models expose controllable reasoning modes and structured tool-call and reasoning parsers. IFM released the TxT360-v2 corpus, specialized code, mathematics, reasoning, SFT, and behavior datasets, the xLLM pretraining codebase, and the Horizon post-training codebase. The exact openness status varies by variant: final weights are available for the 375B and 36B models, the 7B and 3.7B repositories include intermediate checkpoints, and the currently downloadable 32B model is explicitly named K2-Horizon-32B-Stage1 with its final checkpoint still pending. FP8, GGUF, and selected Uno variants are also available. The flagship 375B model reports 70.2 on Terminal-Bench 2.1, 42.6 on SWE-Bench Pro, 72.8 on BrowseComp, 67.7 on MCPMark, 87.3 on GPQA Diamond, and a 1441 GDPval-AA Elo rating. The 7B model reports 70.6 on SWE-Bench Verified, 59.0 on BrowseComp, and 68.0 on AA-LCR, while the 0.9B model reports 48.5 on AIME 2026 and 79.9 on HumanEval+. All benchmark results are vendor-reported and use the evaluation settings and harnesses described in the individual model cards.",
      "paper_url": "",
      "links": [
        "https://ifm.ai/k2/",
        "https://ifm.ai/blog/k2/",
        "https://ifm.ai/k2/press-release/",
        "https://huggingface.co/collections/IFM/k2-horizon",
        "https://huggingface.co/IFM/K2-Horizon-375B-A23B",
        "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B",
        "https://huggingface.co/IFM/K2-Horizon-32B",
        "https://huggingface.co/IFM/K2-Horizon-7B",
        "https://huggingface.co/IFM/K2-Horizon-3.7B",
        "https://huggingface.co/IFM/K2-Horizon-0.9B",
        "https://huggingface.co/datasets/IFM/TxT360-v2",
        "https://github.com/ifm-ai/xllm",
        "https://github.com/ifm-ai/horizon-post-train",
        "https://recipes.vllm.ai/IFM",
        "https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon"
      ],
      "page_url": "https://nlp.com.ai/models/k2-horizon-09b-37b-7b-32b-stage1-36b-a4b-375b-a23b/",
      "record_updated_at": "2026-10-11T08:37:45.806277+00:00"
    },
    {
      "slug": "spark-x25-17b-and-4b",
      "source_file": "2026-09-01-spark-x2.5.json",
      "release_date": "2026-09-01",
      "date": "September 1, 2026",
      "name": "Spark-X2.5 (1.7B and 4B)",
      "developer": "SparkLLM / XHToken (Singapore)",
      "parameters": "1.71B and 4.11B (dense; marketed as 1.7B and 4B)",
      "context_window": "1,048,576 tokens",
      "license": "Apache 2.0",
      "key_features": "Same-day release of two compact, general-purpose dense language models designed for local and on-device conversation, writing, translation, reasoning, coding, tool use, and agentic workflows. Both models use a hybrid attention architecture that repeats three 512-token sliding-window attention layers followed by one full-attention layer, allowing them to natively support a 1,048,576-token context window while reducing KV-cache and compute requirements relative to full attention at every layer. Spark-X2.5-1.7B contains 1.71B parameters, 28 layers, a 2,048-dimensional hidden state, eight attention heads, and two key/value heads; Spark-X2.5-4B contains 4.11B parameters, 36 layers, a 2,560-dimensional hidden state, 16 attention heads, and four key/value heads. Both use 256-dimensional attention heads, partial rotary embeddings, grouped-query attention, headwise attention-output gates, tied embeddings, a 131,072-token vocabulary, and BF16 weights. The family was pretrained on approximately 20 trillion tokens drawn from web pages, books, academic publications, code, and encyclopedic material, followed by a dedicated long-context stage comprising hundreds of billions of tokens at sequence lengths extending to 1M. Post-training combines supervised fine-tuning with large-scale reinforcement learning across language understanding, reasoning, programming, instruction following, and tool-augmented agent behavior; specialized teacher policies are consolidated into the deployable models through MOPD. Thinking is enabled by default but can be disabled through the chat template, and the models support more than 200 languages. SparkLLM provides integrations or deployment guidance for SGLang, vLLM, llama.cpp, MLX, Ollama, LM Studio, LLaMA-Factory, Codex, Claude Code, OpenClaw, Hermes, and DeepSeek Harness. In vendor-reported thinking-mode evaluations, the 4B model scores 65.1 on BFCL-V4, 75.1 on tau2-bench, 30.4 on tau3-bench, 54.6 on MCP-Atlas, 40.9 on BrowseComp, 44.4 on SWE-Bench Pro, 53.3 on SWE-Bench Multilingual, 90.7 on AIME 2026, and 75.0 on IFBench. The 1.7B model scores 46.9 on BFCL-V4, 65.3 on tau2-bench, 29.7 on BrowseComp, 28.3 on SWE-Bench Verified, 69.4 on AIME 2026, and 66.3 on IFBench. Matching Base and GGUF variants are also available through the official release collection.",
      "paper_url": "",
      "links": [
        "https://dev.to/sparkllm/spark-x25-4b-17b-the-only-on-device-models-with-native-1m-token-context-now-open-source-d9o",
        "https://github.com/XHToken/Spark-X2.5",
        "https://huggingface.co/collections/XHToken/spark-x25",
        "https://huggingface.co/XHToken/Spark-X2.5-1.7B",
        "https://huggingface.co/XHToken/Spark-X2.5-4B",
        "https://modelscope.cn/collections/XHToken/Spark-X25",
        "https://modelscope.cn/models/XHToken/Spark-X2.5-1.7B",
        "https://modelscope.cn/models/XHToken/Spark-X2.5-4B"
      ],
      "page_url": "https://nlp.com.ai/models/spark-x25-17b-and-4b/",
      "record_updated_at": "2026-10-11T08:37:45.803178+00:00"
    },
    {
      "slug": "deepseek-v4-flash-vision-exp",
      "source_file": "2026-08-31-deepseek-v4-flash-vision-exp.json",
      "release_date": "2026-08-31",
      "date": "August 31, 2026",
      "name": "DeepSeek-V4-Flash-Vision-Exp",
      "developer": "DeepSeek AI (China)",
      "parameters": "305B total (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Experimental open-weight multimodal continuation of DeepSeek-V4-Flash, released after an August 21 API preview and produced by adding visual modules to the Flash architecture followed by continued training. The official checkpoint contains approximately 305B total parameters across 48 mixed-precision safetensors shards totaling about 168 GB. Its 43-layer sparse MoE language backbone uses a 4,096-dimensional hidden state, 256 routed experts plus one shared expert, six routed experts selected per token, 64 attention heads, one key/value head, DFlash sparse attention with an index top-k of 512 and a 128-token sliding window, Hyper-Connections, and three DSpark next-token-prediction layers for integrated speculative decoding. YaRN scales the original 65,536-token window to 1,048,576 tokens. The added vision tower has 32 layers, a 1,024-dimensional state, 16 attention heads, 14-pixel patches, a downsampling ratio of 3, and a maximum of 384 visual tokens. The repository combines FP4 expert weights with dynamic blockwise FP8 quantization for supported components, provides OpenAI-style multimodal message encoding plus `<image>path</image>` notation, and includes a minimal PyTorch implementation covering the vision encoder and aligner, DFlash, MoE, Hyper-Connections, and DSpark, alongside SGLang deployment guidance. DeepSeek reports text-agent scores of 83.9 on Terminal-Bench 2.1, 59.3 on DeepSWE, 75.9 on Toolathlon-Verified, and 75.3 on CyberGym, plus multimodal scores of 36.5 on ApexBench pass@1, 27.3 on Agents' Last Exam, 64.3 on Chartography, and 35.0 on ZeroBench pass@5. DeepSeek positions it as retaining text-agent performance comparable to DeepSeek-V4-Flash-0731 while substantially improving multimodal-agent capabilities.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
        "https://huggingface.co/collections/deepseek-ai/deepseek-v4",
        "https://api-docs.deepseek.com/news/news260821/",
        "https://api-docs.deepseek.com/guides/vision/",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp/tree/main/encoding",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp/tree/main/inference"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v4-flash-vision-exp/",
      "record_updated_at": "2026-10-11T08:37:45.799879+00:00"
    },
    {
      "slug": "hy4-preview-bf16-fp8",
      "source_file": "2026-08-28-hy4-preview.json",
      "release_date": "2026-08-28",
      "date": "August 28, 2026",
      "name": "Hy4 preview (BF16 + FP8)",
      "developer": "Tencent Hy Team (China)",
      "parameters": "770B total / 49B active MoE backbone + 10B total / 0.7B active MTP layer",
      "context_window": "1,048,576 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight early preview of Tencent's fourth-generation Hy flagship, released as paired BF16 and FP8 instruction-tuned checkpoints. The backbone contains 770 billion total parameters with 49 billion activated per token, alongside one native Multi-Token Prediction layer containing 10B total and 0.7B active parameters for speculative decoding. Its 78-layer architecture uses one dense feed-forward layer followed by 77 Mixture-of-Experts layers, each containing 256 routed experts and one shared expert; every token activates eight routed experts plus the shared expert. The attention system uses Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse-index reuse, while identity Hyper-Connections expand the residual pathway to four streams. Additional configuration details include a 6,144-dimensional hidden state, 64 attention heads, 2,048-dimensional query compression, 512-dimensional key-value compression, a sparse-attention index top-k of 2,048, and a native 1,048,576-token context window. Tencent positions Hy4 preview as a productivity-oriented agent model co-designed with software engineers, game developers, financial analysts, security specialists, CodeBuddy, and WorkBuddy, emphasizing long-horizon software development, frontend interaction quality, cross-file office analysis, financial modeling, playable game-prototype generation, and scientific research. In a blind internal evaluation, 163 Tencent experts rated outputs on 203 engineering tasks; Tencent reports a 2.99 out of 4 average for Hy4 preview versus 2.92 for GLM-5.3 and 2.94 for Kimi K3, with 46.8% wins, 12.8% ties, and 40.4% losses against GLM-5.3 and 51.2% wins, 7.9% ties, and 40.9% losses against Kimi K3. Published maximum-effort evaluations report 82.9 on SWE-bench Multilingual, 65.7 on SWE-bench Pro, 64.3 on DeepSWE, 85.4 on Terminal-Bench 2.1, 78.4 on CyberGym, 74.1 on Toolathlon-Verified, 65.4 on OneMillionBench with tools, 83.7 on MCP-Atlas, 92.3 on GPQA Diamond, 74.2 on MathArena Apex 2025, and 55.4 on Humanity's Last Exam with tools. Reasoning defaults to high effort, while a no_think mode is available for direct responses. Tencent explicitly identifies excessive reasoning time on difficult tasks and a tendency to over-verify its work as known preview limitations. The official Hugging Face repositories contain approximately 1.56 TB of BF16 files across 131 safetensors shards and 814 GB of FP8 files across 130 shards. Official deployment recipes support vLLM and SGLang, MTP speculative decoding, automatic tool selection, and dedicated Hy4 reasoning and tool-call parsers.",
      "paper_url": "",
      "links": [
        "https://hy.tencent.ai/research/hy4-preview",
        "https://github.com/Tencent-Hunyuan/Hy4-preview",
        "https://huggingface.co/tencent/Hy4-preview",
        "https://huggingface.co/tencent/Hy4-preview-FP8",
        "https://modelscope.cn/models/Tencent-Hunyuan/Hy4-preview",
        "https://modelscope.cn/models/Tencent-Hunyuan/Hy4-preview-FP8",
        "https://huggingface.co/tencent/Hy4-preview/raw/main/README.md",
        "https://huggingface.co/tencent/Hy4-preview/raw/main/config.json",
        "https://huggingface.co/tencent/Hy4-preview/raw/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/hy4-preview-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.796731+00:00"
    },
    {
      "slug": "glm-53-fp8",
      "source_file": "2026-08-28-glm-5.3.json",
      "release_date": "2026-08-28",
      "date": "August 28, 2026",
      "name": "GLM-5.3 (FP8)",
      "developer": "Z.AI / Zhipu AI (China)",
      "parameters": "~753B total / 40B active (MoE; same base as GLM-5.2, with the official GLM family specification also described as 744B-A40B)",
      "context_window": "1,048,576 tokens",
      "license": "GLM-5.3 License",
      "key_features": "Official open-weight release of Z.AI's GLM-5.3 flagship, published as a mixed-precision FP8 checkpoint across 141 safetensors shards totaling approximately 756 GB. GLM-5.3 uses the same base model and sparse Mixture-of-Experts architecture as GLM-5.2, with its reported gains coming entirely from expanded post-training on coding, long-horizon engineering, research, tool use, and cybersecurity environments. The official configuration specifies 78 transformer layers, a 6,144-dimensional hidden state, 64 attention heads, 256 routed experts plus one shared expert, eight routed experts selected per token, three initial dense feed-forward layers followed by sparse MoE layers, DeepSeek Sparse Attention with a top-k of 2,048, one Multi-Token Prediction layer, and a native 1,048,576-token context window. The released checkpoint uses dynamic blockwise FP8 quantization with selected modules retained in BF16 or F32. GLM-5.3 is an always-thinking model with low, high, and max reasoning-effort settings, defaulting to max; disabling thinking is no longer supported. Z.AI reports improvements over GLM-5.2 from 4.6 to 28.3 on Terminal-Bench 3.0, 46.2 to 66.9 on DeepSWE v1.1, 19.4 to 42.5 on SWE-Marathon v1.1, 26.2 to 48.2 on AutomationBench, and 77.2 to 84.5 on CyberGym. Official local-deployment guidance covers SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, and Ascend-oriented inference frameworks. The custom GLM-5.3 License broadly permits use, modification, redistribution, derivatives, and commercial deployment, while requiring Z.AI security review for Model-as-a-Service operators whose affiliated annual revenue exceeds 10 billion US dollars.",
      "paper_url": "https://arxiv.org/abs/2602.15763",
      "links": [
        "https://huggingface.co/zai-org/GLM-5.3",
        "https://z.ai/blog/glm-5.3",
        "https://github.com/zai-org/GLM-5",
        "https://arxiv.org/abs/2602.15763",
        "https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/glm-53-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.793666+00:00"
    },
    {
      "slug": "qwen38-flash-next",
      "source_file": "2026-08-26-qwen3.8-flash-next.json",
      "release_date": "2026-08-26",
      "date": "August 26, 2026",
      "name": "Qwen3.8-Flash-Next",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "125B total / 6B active MoE backbone, with 51B N-gram embeddings and 4B MTP parameters",
      "context_window": "262,144 tokens native; extensible to 1,000,000 tokens with YaRN",
      "license": "Qwen Community License 1.0",
      "key_features": "Experimental open-weight multimodal preview of the architecture being developed for Qwen4, distinct from the forthcoming production Qwen3.8-Flash managed service. The post-trained model accepts text, images, and video and produces text, using a 125-billion-parameter sparse backbone with approximately six billion parameters activated per token, 512 experts with ten routed experts plus one shared expert selected per token, 51 billion parameters of host-resident bigram and trigram embeddings, and four billion Multi-Token Prediction parameters. Its 48 language-model layers comprise 12 repetitions of three Gated DeltaNet-plus-MoE layers followed by one Qwen Sparse Attention-plus-MoE layer. Additional architectural experiments include four-branch Gated Residual connections, a 20-million-entry N-gram embedding table inserted at layer two, QSA retrieval over 2,048 selected tokens or 512 blocks, and Muon optimization for two-dimensional linear parameters. The technical report attributes an approximately ninefold reduction in training FLOPs relative to its 397B-A17B predecessor to using roughly one-third as many active parameters and one-third as many training tokens; it also reports that QSA's attention kernel is 7.6 times faster during one-million-token prefill and 4.9 times faster during decoding than dense attention. The model supports thinking and non-thinking operation, xhigh, medium, and low reasoning-effort levels, and preserve_thinking enabled by default so that historical reasoning can be retained across multi-turn and tool-calling workflows. Published evaluations report 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual, 73.9 on CoWorkBench, 91.7 on GPQA Diamond, 91.9 on LiveCodeBench v6, 84.5 on AndroidWorld, 64.0 on Vision2Web, and 95.7 on MathVision with computer interaction. The approximately 360 GB release is distributed across 131 BF16 safetensors shards. Qwen separately announced that the forthcoming production Qwen3.8-Flash service will provide default one-million-token context, built-in tools, and pricing of $0.16 per million input tokens and $0.47 per million output tokens. The Qwen Community License requires prominent Qwen attribution for products exceeding 100 million monthly active users or $20 million in monthly revenue and requires a separate commercial license for third-party Model-as-a-Service or AI work-assistant businesses, excluding internal use.",
      "paper_url": "https://github.com/QwenLM/Qwen3.8-Flash-Next/blob/main/tech_report.pdf",
      "links": [
        "https://qwen.ai/blog?id=qwen3.8-flash-next",
        "https://github.com/QwenLM/Qwen3.8-Flash-Next",
        "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
        "https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next",
        "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/raw/main/README.md",
        "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/raw/main/config.json",
        "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/raw/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/qwen38-flash-next/",
      "record_updated_at": "2026-10-11T08:37:45.790543+00:00"
    },
    {
      "slug": "glm-53-flash-fp8-bf16",
      "source_file": "2026-08-26-glm-5.3-flash.json",
      "release_date": "2026-08-26",
      "date": "August 26, 2026",
      "name": "GLM-5.3-Flash (FP8, BF16)",
      "developer": "Z.AI / Zhipu AI (China)",
      "parameters": "320B total / 18B active (MoE; official checkpoint contains approximately 321.3B parameters)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Official open-weight release of GLM-5.3-Flash, comprising native FP8 and BF16 checkpoints and confirming the identity of the anonymously tested ox-alpha stealth model. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series and is distinct from the full GLM-5.3 flagship: whereas GLM-5.3 continues post-training from the GLM-5.2 base, Flash begins from a newly trained base model with a redesigned architecture and a 30-trillion-token multimodal pre-training corpus. The model contains approximately 320 billion total parameters with 18 billion activated per token, versus 744B-A40B for GLM-5.2. Its 45-layer, 4,096-dimensional language backbone generally follows a repeating pattern of three linear-attention layers followed by one sparse-attention layer, with 34 linear and 11 sparse layers in total. The first three feed-forward layers are dense; subsequent layers use 288 routed experts with eight selected per token plus one shared expert. Architectural additions include four-branch Manifold-Constrained Hyper-Connections, one Multi-Token Prediction layer, and IndexPool, which compresses groups of four sparse-attention indexer keys through weighted pooling. Z.AI reports that the hybrid linear-and-sparse attention design reduces attention computation by 3.0 times and average KV-cache size by 4.4 times relative to the full GLM-5.3 at one-million-token context. Native visual processing is provided by a 24-layer, 1,024-dimensional vision encoder with temporal patching, supporting visual coding, document understanding, browser use, computer use, output inspection, and self-verification. The embedded chat template supports low, high, and max reasoning-effort settings, defaults to max, and clears historical reasoning traces unless clear_thinking is set to false. Published evaluations report 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE v1.1, 56.3 on NL2Repo, 78.4 on Toolathlon Verified, 48.8 on AutomationBench, 26.3 on Agents' Last Exam, and 55.3 on Humanity's Last Exam with tools. Multimodal evaluations include 62.4 on OfficeQA Pro, 89.4 on CharXiv Reasoning with tools, 78.0 on Chartography with tools, 53.4 on BabyVision, 77.8 on MVBench, and 80.5 on MMVU. Z.AI also reports an Artificial Analysis Intelligence Index v4.1.1 score of 57 at a discounted cost of $0.045 per task, describing the model as outperforming GLM-5.2 across benchmarks and real-world workloads at one-tenth the price while approaching Claude Opus 4.8 on coding and agentic benchmarks. The primary FP8 repository contains approximately 314.4 billion FP8 and 6.9 billion BF16 parameters across 62 shards, while the full-BF16 repository contains approximately 321.3 billion BF16 parameters across 120 shards. Local deployment guidance is available for SGLang, vLLM, TokenSpeed, and KTransformers.",
      "paper_url": "https://arxiv.org/abs/2602.15763",
      "links": [
        "https://z.ai/blog/glm-5.3-flash",
        "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "https://huggingface.co/zai-org/GLM-5.3-Flash-BF16",
        "https://modelscope.cn/models/ZhipuAI/GLM-5.3-Flash",
        "https://modelscope.cn/models/ZhipuAI/GLM-5.3-Flash-BF16",
        "https://github.com/zai-org/GLM-5",
        "https://docs.z.ai/guides/llm/glm-5.3-flash",
        "https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/README.md",
        "https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/config.json",
        "https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/glm-53-flash-fp8-bf16/",
      "record_updated_at": "2026-10-11T08:37:45.787265+00:00"
    },
    {
      "slug": "granite-42-3b-8b-30b",
      "source_file": "2026-08-25-granite-4.2.json",
      "release_date": "2026-08-25",
      "date": "August 25, 2026",
      "name": "Granite 4.2 (3B, 8B, 30B)",
      "developer": "IBM (USA)",
      "parameters": "3.66B, 8.79B, 29.28B dense (marketed as 3B, 8B, 30B)",
      "context_window": "131,072 tokens in released configs; IBM describes long-context extension to 512K",
      "license": "Apache License 2.0",
      "key_features": "Official open-weight reasoning-focused Granite release spanning three dense decoder-only models marketed as 3B, 8B, and 30B. IBM's Hugging Face collection reports exact parameter counts of approximately 3.66B, 8.79B, and 29.28B, while the released GraniteForCausalLM configurations use grouped-query attention, SwiGLU feed-forward layers, rotary position embeddings, and 131,072 maximum position embeddings. The official repository describes Granite 4.2 as post-trained on Granite 4.1 base checkpoints, while IBM's technical blog describes an underlying five-phase pretraining program of approximately 15 trillion tokens culminating in 512K long-context training; the published model cards consequently characterize 128K as native context with extension to 512K. All three models provide explicit chain-of-thought reasoning with full thinking, non-thinking, and low-effort modes, plus native OpenAI-compatible tool calling. Previous assistant reasoning is stripped from multi-turn history by default to conserve context, but the chat template exposes truncate_history_thinking=False for applications that want to preserve complete reasoning history. Supervised fine-tuning uses approximately 7.2 million samples and 100 billion tokens spanning reasoning, code, math, multilingual, safety, and agentic trajectories, followed by multi-environment GRPO and RLHF. The 8B and 30B models receive additional agentic reinforcement learning across sandboxed software-engineering, terminal, tool-use, and web-search environments. IBM reports 89.17 on AIME25, 75.77 on LiveCodeBench v6, 57.0 on SWE-bench Verified, 33.29 on SWE-bench Pro, 29.24 on Terminal-Bench 2.1, and 81.38 on RULER 128K for the 30B model. Official deployment artifacts include BF16 weights and IBM-produced GGUF, FP8, NVFP4, and MXFP4 quantizations, with serving guidance for Transformers, vLLM, SGLang, OpenCode, Pi, and OpenHands.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/blog/ibm-granite/granite-4-2",
        "https://github.com/ibm-granite/granite-4.2-language-models",
        "https://huggingface.co/collections/ibm-granite/granite-42-language-models",
        "https://huggingface.co/ibm-granite/granite-4.2-3b",
        "https://huggingface.co/ibm-granite/granite-4.2-8b",
        "https://huggingface.co/ibm-granite/granite-4.2-30b"
      ],
      "page_url": "https://nlp.com.ai/models/granite-42-3b-8b-30b/",
      "record_updated_at": "2026-10-11T08:37:45.784214+00:00"
    },
    {
      "slug": "thomson-10-small",
      "source_file": "2026-08-24-thomson-1.0-small.json",
      "release_date": "2026-08-24",
      "date": "August 24, 2026",
      "name": "Thomson-1.0-Small",
      "developer": "Thomson Reuters (Canada)",
      "parameters": "35.107B total / approximately 3B active (MoE)",
      "context_window": "262,144 tokens natively",
      "license": "PolyForm Strict License 1.0.0 (noncommercial and restricted-use)",
      "key_features": "Official open-weight member of the Thomson-1.0 family, developed for specialized and general-purpose work with an emphasis on high-stakes professional domains including legal, tax, and journalism. Thomson-1.0-Small is a multimodal mixture-of-experts model derived from Qwen3.6-35B-A3B through the intermediate Snowdon1.1-Small checkpoint, with 35.107B total parameters and approximately 3B activated per token. Its Qwen3.5-MoE conditional-generation configuration retains image and video processing, uses 40 language layers, 256 routed experts with top-8 activation, hybrid linear and full attention, and a native 262,144-token context window. The development pipeline combines constitutional value re-alignment, continual pre-training, Direct Preference Optimisation, and reinforcement learning. Continual pre-training used 200B tokens curated from more than 19T tokens of permissively public and Thomson Reuters proprietary material, including news, contracts, regulatory filings, case law, statutes, and practitioner guidance, with replay data intended to preserve general capabilities. Post-training includes document- and ontology-driven preference data, practitioner queries, and agentic Deep Research training emphasizing faithful tool use and citation behavior. The model supports reasoning, structured tool calls, text generation, and image/video inputs through its supplied chat template and processor configuration. Thomson Reuters reports a 74.6 cross-domain benchmark average, including domain averages of 75.2 for legal and 82.6 for tax, alongside improvements over its Qwen3.6 base on document processing, human professional queries, deep research, factuality, and general-agent evaluations. Training consumed approximately 1.63e23 FLOP and 35,207 NVIDIA B200 GPU-hours. The weights are released under PolyForm Strict 1.0.0, which permits specified noncommercial uses but restricts commercial use, redistribution, modification, and derivative works.",
      "paper_url": "https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Thomson_1_0_Technical_Report.pdf",
      "links": [
        "https://www.thomsonreuters.com/en/press-releases/2026/august/thomson-reuters-leverages-its-world-class-data-assets-to-launch-its-own-frontier-model",
        "https://www.thomsonreuters.com/en/thomson-llm",
        "https://huggingface.co/thomsonreuters/Thomson-1.0-Small",
        "https://huggingface.co/thomsonreuters/Thomson-1.0-Small/blob/main/config.json",
        "https://huggingface.co/thomsonreuters/Thomson-1.0-Small/blob/main/LICENSE",
        "https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Thomson_1_0_Technical_Report.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/thomson-10-small/",
      "record_updated_at": "2026-10-11T08:37:45.781046+00:00"
    },
    {
      "slug": "ornith-15",
      "source_file": "2026-08-19-ornith-1.5.json",
      "release_date": "2026-08-19",
      "date": "August 19, 2026",
      "name": "Ornith-1.5",
      "developer": "Ornith Team",
      "parameters": "9B dense; 35B total / approximately 3B active MoE; 397B MoE",
      "context_window": "262,144 tokens natively (approximately 1M with validated YaRN scaling)",
      "license": "MIT",
      "key_features": "MIT-licensed open-weight release of the Ornith-1.5 family across 9B dense, 35B-A3B mixture-of-experts, and flagship 397B mixture-of-experts scales. The released checkpoints use Qwen3.5 dense and MoE conditional-generation architectures, with official configurations retaining Qwen3.5 vision encoders and image and video token support. Ornith-1.5 expands the self-scaffolding reinforcement-learning framework introduced with Ornith-1.0 into a broader self-improvement loop in which the model proposes new training tasks, generates or refines task-specific scaffolds and evaluation harnesses, produces solution rollouts, and jointly optimizes all three stages with GRPO. Generated tasks are rewarded according to validity and verifiability, proximity to the model's current capability frontier, and novelty relative to previously generated tasks, while harnesses are evaluated for task alignment, reward fidelity, and resistance to reward hacking. All three checkpoints are reasoning models that emit separate thinking traces and structured tool calls suitable for OpenAI-compatible APIs and agent frameworks. They provide 262,144-token native context windows and support tested YaRN scaling to approximately one million tokens, although the team cautions that static scaling can reduce quality on ordinary-length requests. The official Hugging Face collection provides BF16 checkpoints for all three scales, GGUF versions for all three, FP8 and NVFP4 variants for the two MoE models, and MLX variants for the 9B and 35B models. Self-reported results averaged over five runs include 86.1 on Terminal-Bench 2.1, 86.0 on SWE-bench Verified, and 56.0 on DeepSWE for Ornith-1.5-397B; 68.5, 79.0, and 22.0 respectively for Ornith-1.5-35B-A3B; and 47.0 on Terminal-Bench 2.1 plus 70.6 on SWE-bench Verified for Ornith-1.5-9B. The evaluation methodology includes network isolation, removal of Git history, and other anti-hacking measures on several software-engineering benchmarks.",
      "paper_url": "",
      "links": [
        "https://ornith.ai/ornith_1_5.html",
        "https://huggingface.co/collections/ornith-ai/ornith-15",
        "https://huggingface.co/ornith-ai/Ornith-1.5-9B",
        "https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B",
        "https://huggingface.co/ornith-ai/Ornith-1.5-397B"
      ],
      "page_url": "https://nlp.com.ai/models/ornith-15/",
      "record_updated_at": "2026-10-11T08:37:45.777467+00:00"
    },
    {
      "slug": "ling-30-base-checkpoints-tiny-and-flash",
      "source_file": "2026-08-19-ling-3.0-base-checkpoints.json",
      "release_date": "2026-08-19",
      "date": "August 19, 2026",
      "name": "Ling-3.0 Base Checkpoints (Tiny and Flash)",
      "developer": "inclusionAI (China)",
      "parameters": "7.9B total / 1.3B active (Tiny); 124B total / 5.1B active (Flash)",
      "context_window": "262,144 tokens",
      "license": "MIT",
      "key_features": "Research-oriented release of six unaligned Ling-3.0 base-model checkpoints spanning two sparse Mixture-of-Experts scales and three stages of the same training recipe. For both Ling-3.0-tiny and Ling-3.0-flash, inclusionAI provides a `-base-30T` checkpoint taken after large-scale pretraining but before mid-training, learning-rate decay or WSM merging, and post-training; a `-base-midtrain` checkpoint taken after mid-training but before merging or post-training; and an unsuffixed `-base` checkpoint produced from the mid-trained model through Warmup-Stable and Merge, or WSM. WSM replaces a conventional learning-rate decay phase with weighted checkpoint merging, making the resulting models more suitable for continual pretraining and allowing researchers to explore different effective decay profiles offline without rerunning the corresponding training schedules. Tiny and Flash share one training recipe so that continued-pretraining, fine-tuning, and post-training strategies can be developed on the smaller model before being scaled to Flash. Ling-3.0-tiny contains 7.9B total and 1.3B active parameters across 24 transformer layers arranged as 18 Kimi Delta Attention and 6 gated Multi-Head Latent Attention layers in a 3:1 pattern. It has one initial dense layer, 128 routed experts, and activates eight routed experts plus one shared expert per token. Ling-3.0-flash contains 124B total and 5.1B active non-embedding parameters across 42 layers arranged as 35 Kimi Delta Attention and 7 gated Multi-Head Latent Attention layers in a 5:1 pattern. It has two initial dense layers, 512 routed experts, and activates eight routed experts plus one shared expert per token. Both architectures use native hybrid linear attention, a 157,184-token vocabulary, and configurations supporting 262,144-token contexts. None of the six checkpoints has undergone supervised or reinforcement-learning post-training; inclusionAI recommends them for continued pretraining, mid-training, supervised fine-tuning, preference optimization, RL, distillation, and long-context or MoE research rather than direct end-user deployment.",
      "paper_url": "https://arxiv.org/abs/2507.17634",
      "links": [
        "https://huggingface.co/collections/inclusionAI/ling-30",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny-base-30T",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny-base-30T",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny-base-midtrain",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny-base-midtrain",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny-base",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-tiny-base",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-base-30T",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-base-30T",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-base-midtrain",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-base-midtrain",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-base",
        "https://modelscope.cn/models/inclusionAI/Ling-3.0-flash-base"
      ],
      "page_url": "https://nlp.com.ai/models/ling-30-base-checkpoints-tiny-and-flash/",
      "record_updated_at": "2026-10-11T08:37:45.774388+00:00"
    },
    {
      "slug": "qwen38-27b",
      "source_file": "2026-08-14-qwen3.8-27b.json",
      "release_date": "2026-08-14",
      "date": "August 14, 2026",
      "name": "Qwen3.8-27B",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "27B dense language model; 27,781,427,952 total BF16 parameters including vision encoder",
      "context_window": "262,144 tokens native; extensible to 1,000,000 tokens with YaRN",
      "license": "Apache License 2.0",
      "key_features": "Compact open-weight dense model from the Qwen3.8 generation, released as a native vision-language model supporting text, image, and video input with text output. Qwen3.8-27B builds on the Qwen3.5 hybrid architecture with a 5,120-dimensional hidden state and 64 language-model layers arranged as 16 repetitions of three Gated DeltaNet-plus-FFN blocks followed by one gated-attention-plus-FFN block, supplemented by Multi-Token Prediction training. The model supports 262,144 tokens natively and extension to one million tokens through YaRN. Unlike the open Qwen3.8-2.4T-A95B checkpoint, the 27B release retains native vision and video understanding and permits thinking to be disabled per request. Thinking is enabled by default, with xhigh, medium, and low reasoning-effort levels, while preserve_thinking can retain reasoning traces across multi-turn agent workflows. Qwen positions the model for coding, professional work, research, long-horizon agents, autonomous planning, environment-feedback handling, computer interaction, and broad agent-harness compatibility. Published evaluations report 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro, 42.2 on DeepSWE 1.1, 79.0 on QwenSWEBench, 70.7 on CoWorkBench, 79.5 on IFBench, and 89.2 on GPQA Diamond. The approximately 55.6 GB BF16 release is distributed across 18 safetensors shards with official deployment guidance for Transformers, vLLM, SGLang, and TokenSpeed.",
      "paper_url": "",
      "links": [
        "https://qwen.ai/blog?id=qwen3.8",
        "https://huggingface.co/Qwen/Qwen3.8-27B",
        "https://modelscope.cn/models/Qwen/Qwen3.8-27B",
        "https://huggingface.co/Qwen/Qwen3.8-27B/raw/main/README.md",
        "https://huggingface.co/Qwen/Qwen3.8-27B/raw/main/config.json",
        "https://huggingface.co/Qwen/Qwen3.8-27B/raw/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/qwen38-27b/",
      "record_updated_at": "2026-10-11T08:37:45.771162+00:00"
    },
    {
      "slug": "intern-s2-mobius-bf16-fp8",
      "source_file": "2026-08-14-intern-s2-mobius.json",
      "release_date": "2026-08-14",
      "date": "August 14, 2026",
      "name": "Intern-S2-Mobius (BF16, FP8)",
      "developer": "Shanghai AI Laboratory / InternLM (China)",
      "parameters": "35B-class sparse multimodal model; 35,964,405,616 parameters in the BF16 checkpoint",
      "context_window": "262,144 tokens",
      "license": "Apache-2.0",
      "key_features": "Official open-weight scientific multimodal foundation model built on InternLM’s experimental Mobius-v0 architecture and continued-pretrained from Qwen3.5-35B before supervised fine-tuning and reinforcement learning. Rather than binding feed-forward knowledge storage to individual Transformer layers, Mobius separates knowledge from reasoning computation through a globally shared Memory containing 2,560 small routed experts and multiple recurrent Reasoners that repeatedly query and refine hidden states against that shared repository. This design enables Backward Residual Connection, allowing reasoning stages to access knowledge outside their local forward layer hierarchy, and Dynamic Latent Reasoning, which performs iterative deliberation and refinement in continuous hidden states before decoding. The published configuration uses 40 language-model layers with a three-to-one mixture of linear-attention and full-attention layers, 16 attention heads, eight selected memory experts per token, a 2,048-dimensional hidden state, and a separate Multi-Token Prediction layer with 256 experts and eight selected per token. InternLM reports that Mobius matches or improves upon Qwen3.5-35B across aggregate general and scientific reasoning evaluations while producing substantially shorter visible reasoning traces, yielding nearly four times higher end-to-end inference efficiency in its reported evaluation. Scientific gains are highlighted on Biology-Instructions, Mol-Instructions, and MolecularIQ. The checkpoint supports text, image, video, scientific sequence, and time-series inputs; native thinking traces and structured tool calling are exposed through its chat template. BF16 and FP8 weights are released under Apache 2.0, with deployment guidance for Transformers, LMDeploy, vLLM, and SGLang and recommended MTP speculative decoding support.",
      "paper_url": "https://github.com/InternLM/Intern-S2-Mobius/blob/main/Technical_Report_Intern_S2_Mobius.pdf",
      "links": [
        "https://huggingface.co/internlm/Intern-S2-Mobius",
        "https://huggingface.co/internlm/Intern-S2-Mobius-FP8",
        "https://huggingface.co/collections/internlm/intern-s2",
        "https://github.com/InternLM/Intern-S2-Mobius",
        "https://github.com/InternLM/archspace"
      ],
      "page_url": "https://nlp.com.ai/models/intern-s2-mobius-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.767840+00:00"
    },
    {
      "slug": "dots3-note-preview",
      "source_file": "2026-08-13-dots3-note-preview.json",
      "release_date": "2026-08-13",
      "date": "August 13, 2026",
      "name": "dots3-note Preview",
      "developer": "dots studio / Xiaohongshu (China)",
      "parameters": "280B total / 16B active (multimodal MoE)",
      "context_window": "524,288 tokens",
      "license": "Apache License 2.0",
      "key_features": "First open-weight model in Xiaohongshu's dots3 family and its most lightweight member, designed as a natively multimodal general-purpose and agentic model accepting text, images, video, and audio while producing text. Its language backbone has one dense layer followed by 45 MoE layers, a 5,120-dimensional hidden state, 256 routed experts plus one shared expert with top-8 routing, 13 DSA attention layers, 33 sliding-window-attention layers, and a 152K vocabulary. The multimodal stack adds a sparse MoE vision transformer with 7B total and 1.2B activated parameters plus a dense 800M-parameter audio encoder. Official positioning emphasizes mathematical and logical reasoning, code generation, tool use, multi-step agent workflows, long-context processing, and interactive tasks involving exploration, memory updates, and adaptation. The chat template supports switchable thinking through enable_thinking, while native structured tool calls are supported through the dots tool-call format. Published headline evaluations report 73.4 on Claw-Eval, 61.0 on SWE-bench Pro, 83.3 on BrowseComp, 81.4 on ARC-AGI-2, 75.1 on Terminal-Bench 2.1, and a 3056 Codeforces rating. Multimodal claims include 83.1 on CharXiv-RQ, 72.5 on SimpleVQA, 60.7 on GDPval PDF, 39.3 on MMEVideo v2, and 69.8 on MMAU-Pro. BF16 and FP8 checkpoints are provided, with the FP8 release recommended for one eight-GPU node. Native vLLM support is available on main, while Transformers and SGLang integration pull requests remain under review. The architecture includes a shared 1.13B-parameter Multi-Token Prediction layer; optional MTP/NEXTN speculative decoding is claimed to reduce time per output token by more than 50 percent.",
      "paper_url": "",
      "links": [
        "https://studio.dots.ai/dots/dots3-en.html",
        "https://github.com/studio-dots-ai/dots3-note-prev",
        "https://huggingface.co/dots-studio/dots3-note-prev",
        "https://modelscope.cn/models/dots-studio/dots3-note-prev",
        "https://huggingface.co/dots-studio/dots3-note-prev-fp8",
        "https://modelscope.cn/models/dots-studio/dots3-note-prev-fp8",
        "https://huggingface.co/dots-studio/dots3-note-prev/raw/main/README.md",
        "https://huggingface.co/dots-studio/dots3-note-prev/raw/main/config.json"
      ],
      "page_url": "https://nlp.com.ai/models/dots3-note-preview/",
      "record_updated_at": "2026-10-11T08:37:45.764839+00:00"
    },
    {
      "slug": "deepseek-v4-pro-0813",
      "source_file": "2026-08-13-deepseek-v4-pro-0813.json",
      "release_date": "2026-08-13",
      "date": "August 13, 2026",
      "name": "DeepSeek-V4-Pro-0813",
      "developer": "DeepSeek AI (China)",
      "parameters": "1.65T checkpoint / 49B active target model (MoE; includes attached DSpark module)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Official DeepSeek-V4-Pro release superseding the April preview, with substantially enhanced agentic capabilities and especially large reported gains on production-oriented tasks. The checkpoint builds on the V4-Pro target architecture and bundles an attached DSpark speculative-decoding model in the same repository. Its published configuration describes a 61-layer DeepSeek-V4 causal architecture with 384 routed experts plus one shared expert, six routed experts selected per token, a 7,168-dimensional hidden state, 128 query-attention heads, one key/value head, compressed sparse-attention components, FP4 expert weights, FP8 quantization for other supported components, and YaRN scaling from an original 65,536-token window to 1,048,576 tokens. DeepSeek reports 87.9 on Terminal Bench 2.1, 61.5 on NL2Repo, 83.3 on Cybergym, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, 25.7 on Agents' Last Exam, and 31.8 on AutomationBench Public. The model also reaches 42.7 on Humanity's Last Exam without tools and 60.0 with tools, plus 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard in DeepSeek's internal evaluations. It supports low, high, and max reasoning-effort settings, uses DeepSeek's custom OpenAI-compatible message encoding rather than a Jinja chat template, and includes local deployment support for vLLM, SGLang, and DeepSeek's reference inference implementation. DeepSeek recommends temperature 1.0 and top-p 0.95 for agent tasks and permits output limits up to 384K tokens for high and max reasoning effort.",
      "paper_url": "https://arxiv.org/abs/2606.19348",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813",
        "https://api-docs.deepseek.com/updates/",
        "https://arxiv.org/abs/2606.19348",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/tree/main/encoding",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/tree/main/inference"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v4-pro-0813/",
      "record_updated_at": "2026-10-11T08:37:45.761796+00:00"
    },
    {
      "slug": "qwen38-24t-a95b",
      "source_file": "2026-08-12-qwen3.8-2.4t-a95b.json",
      "release_date": "2026-08-12",
      "date": "August 12, 2026",
      "name": "Qwen3.8-2.4T-A95B",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "2.4T total / 95B active (MoE)",
      "context_window": "262,144 tokens native; extensible to 1,010,000 tokens",
      "license": "Qwen3.8-Max License",
      "key_features": "Qwen's first open-weight Max-class release and the largest model in the Qwen open family at launch, provided as a post-trained causal language model with 2.4 trillion total parameters and 95 billion activated per token. Built on the Qwen3.5 architectural foundation, the model has 92 layers arranged as 23 repetitions of three Gated DeltaNet-plus-MoE blocks followed by one gated-attention-plus-MoE block. Its sparse architecture contains 512 experts with ten routed experts and one shared expert active per token, a hidden dimension of 8,192, a padded 248,320-token vocabulary, grouped-query attention with 64 query and four key-value heads, and multi-step Multi-Token Prediction training. Qwen positions the release for coding, professional work, research, and long-horizon agentic execution, emphasizing stronger autonomous planning, environment-feedback handling, harness compatibility, and reliable end-to-end completion of complex tasks. The checkpoint supports adjustable reasoning depth through xhigh, medium, and low reasoning_effort levels and retains historical reasoning using preserve_thinking. It is text-only and requires thinking mode: every response begins with reasoning enclosed in think tags, and non-thinking generation cannot be enabled. This distinguishes the downloadable checkpoint from the hosted Qwen3.8-Max service, which is based on it but adds vision input, non-thinking support, built-in tools, and a default one-million-token context configuration. Published Qwen evaluations report 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 73.5 on FrontierSWE, 93.0 on PaperBench, 81.9 on WideSearch, 92.6 on GPQA Diamond, 82.8 on IFBench, and 92.9 on MRCR v2 at 256K. The approximately 4.89 TB release is distributed in Hugging Face Transformers format with deployment support documented for SGLang, vLLM, and TokenSpeed.",
      "paper_url": "",
      "links": [
        "https://qwen.ai/blog?id=qwen3.8",
        "https://modelscope.cn/models/Qwen/Qwen3.8-2.4T-A95B",
        "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B",
        "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/README.md",
        "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/config.json",
        "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE",
        "https://www.qwencloud.com/models/qwen3.8-max"
      ],
      "page_url": "https://nlp.com.ai/models/qwen38-24t-a95b/",
      "record_updated_at": "2026-10-11T08:37:45.758851+00:00"
    },
    {
      "slug": "north-micro-vision-instruct",
      "source_file": "2026-08-12-north-micro-vision-instruct.json",
      "release_date": "2026-08-12",
      "date": "August 12, 2026",
      "name": "North Micro Vision Instruct",
      "developer": "Cohere / Cohere Labs (Canada)",
      "parameters": "2.4B total (2B language model + 400M vision encoder)",
      "context_window": "128K text backbone; multimodal training and validated operation up to 8K",
      "license": "Apache 2.0",
      "key_features": "Compact open-weight vision-language model built for prototyping, task-specific fine-tuning, and specialized multimodal applications. North Micro Vision uses Cohere's CohereCompass architecture: a 28-layer, 2B language backbone interleaving 4K sliding-window attention and full attention in a 3:1 schedule, paired with a custom-trained 400M vision encoder initialized from SigLIP 2 SO400M. It processes native-resolution images without forcing a fixed aspect ratio, accepts interleaved and multiple images, and supports visual question answering, captioning, OCR, document and chart understanding, multilingual image understanding, object counting, and grounding with normalized 0-1000 bounding boxes. It has a 262,144-token vocabulary and supports English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese, Arabic, and additional languages. Official evaluations report 0.921 on DocVQA, 0.808 on ChartQA, 0.792 on OCRBench, 0.775 on AI2D, 0.732 average on RefCOCO, 0.725 on CountBench, and 0.622 on RealWorldQA. The release is distributed as a single approximately 5 GB BF16 safetensors checkpoint with Transformers support. It is not presented as a reasoning, coding, tool-calling, or agentic model; system prompts are not recommended because they were not used during training, and multimodal contexts beyond 8K rely on unbenchmarked extrapolation.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct",
        "https://huggingface.co/CohereLabs/North-Micro-Vision-Instruct",
        "https://huggingface.co/CohereLabs/North-Micro-Vision-Instruct/raw/main/README.md",
        "https://huggingface.co/CohereLabs/North-Micro-Vision-Instruct/raw/main/config.json"
      ],
      "page_url": "https://nlp.com.ai/models/north-micro-vision-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.755780+00:00"
    },
    {
      "slug": "lfm25-vl-3b",
      "source_file": "2026-08-12-lfm2.5-vl-3b.json",
      "release_date": "2026-08-12",
      "date": "August 12, 2026",
      "name": "LFM2.5-VL-3B",
      "developer": "Liquid AI (USA)",
      "parameters": "3.1B total (LFM2.5-2.6B language backbone + 400M SigLIP2 NaFlex vision encoder)",
      "context_window": "32,768 tokens",
      "license": "LFM Open License v1.0",
      "key_features": "Open-weight vision-language model built through further mid- and post-training of LFM2-VL-3B, combining the LFM2.5-2.6B language backbone with a shape-optimized 400M-parameter SigLIP2 NaFlex vision encoder. The model processes text and one or more images at native resolution, splitting large images into non-overlapping 512x512 patches alongside a resized whole-image thumbnail, and supports 16 languages with a 128,000-token vocabulary. Liquid AI positions the release for direct, single-turn, high-throughput visual tasks, with improved natural-language object grounding, ScreenSpot-style GUI understanding, multi-image comprehension, and full-page OCR that emits experimental structured layout annotations with normalized bounding boxes, OTSL tables, and LaTeX equations. It also supports Pythonic structured function calls through dedicated tool-call tokens. Compared with LFM2-VL-3B, published results improve from 57.1 to 87.9 on RefCOCO, 50.2 to 61.5 on BLINK, 34.9 to 58.3 on MuirBench, 26.4 to 59.5 on ToolSandBox, 20.5 to 32.5 on BFCLv4, and 43.9 to 47.5 on the English subset of OCRBenchv2, while reaching 80.7 on ScreenSpot-v2. Liquid reports decoding speeds of 228 tokens per second on an Apple M5 Max, 116 tokens per second on an AMD Ryzen AI Max+ 395, and 20 tokens per second on a Galaxy S26 Ultra while fitting in about 3 GB of memory. On one H100 with vLLM, the model reportedly reaches approximately 11,000 output tokens per second at high concurrency and about 34 milliseconds time to first token for a five-frame clip. Official native, GGUF, ONNX, and MLX releases support Transformers, vLLM, SGLang, llama.cpp, WebGPU, and Apple Silicon deployment. Liquid does not recommend the model for long-context or reasoning-intensive visual tasks.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://www.liquid.ai/blog/lfm2-5-vl-3b",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-3B",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-3B-GGUF",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-3B-ONNX",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-3B-MLX-8bit",
        "https://arxiv.org/abs/2511.23404"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-vl-3b/",
      "record_updated_at": "2026-10-11T08:37:45.753014+00:00"
    },
    {
      "slug": "nvidia-nemotron-35-lightning-30b-a3b",
      "source_file": "2026-08-11-nemotron-3.5-lightning-30b-a3b.json",
      "release_date": "2026-08-11",
      "date": "August 11, 2026",
      "name": "NVIDIA Nemotron 3.5 Lightning 30B-A3B",
      "developer": "NVIDIA Corporation (USA)",
      "parameters": "30B total / 3B active (MoE)",
      "context_window": "1,048,576 tokens (1M; 256K used for single-H100 BF16 deployment)",
      "license": "OpenMDW License Agreement, version 1.1",
      "key_features": "Open-weight general-purpose reasoning and agentic model built on NVIDIA's Nemotron-3-Lightning architecture. Nemotron 3.5 Lightning uses a sparse hybrid Mixture-of-Experts design with interleaved Mamba-2, MoE, and selected attention layers, activating 3B of its 30B parameters per token. It was pretrained on more than 20 trillion tokens and includes Multi-Token Prediction heads; subsequent supervised training covered code, mathematics, science, tool calling, structured output, instruction following, long-range retrieval, and multi-document aggregation, followed by asynchronous multi-environment GRPO reinforcement learning across reasoning, multi-step tool use, multi-turn conversation, and coding environments. The model supports up to a 1M-token context window, automatic schema-based tool calling, switchable thinking and non-thinking modes, a runtime reasoning budget, and English, Spanish, French, German, Italian, Japanese, and programming languages. NVIDIA positions the optimized release for long-running autonomous agents, local subagent deployments, chatbots, and RAG systems. Official variants include full-precision BF16 reference and base weights plus an NVFP4 checkpoint that also supports W4A16 serving on Ampere hardware; a GGUF conversion is published through ggml-org. The release supports vLLM, SGLang, TensorRT-LLM, and speculative decoding through native MTP plus external DSpark semi-autoregressive and DFlash block-diffusion draft models. NVIDIA reports BF16 scores of 81.94 on MMLU-Pro, 75.44 on GPQA Diamond, 51.56 on SWE-Bench Verified, 36.97 on BrowseComp, 85.37 on PinchBench, and 832 Elo on GDPval-AA-V2; the NVFP4 variant scores 52.80 on SWE-Bench Verified, 36.81 on BrowseComp, and 865 Elo on GDPval-AA-V2. NVIDIA also published reproducible NeMo Gym and NeMo Evaluator recipes for its evaluation suite.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16",
        "https://huggingface.co/ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
        "https://github.com/NVIDIA-NeMo/Gym/blob/main/nemotron_recipes/lightning-3.5/reproducibility.md",
        "https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.1/LICENSE.OpenMDW-1.1"
      ],
      "page_url": "https://nlp.com.ai/models/nvidia-nemotron-35-lightning-30b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.750504+00:00"
    },
    {
      "slug": "muse-glimmer-30b",
      "source_file": "2026-08-10-muse-glimmer-30b.json",
      "release_date": "2026-08-10",
      "date": "August 10, 2026",
      "name": "Muse Glimmer 30B",
      "developer": "Meta Superintelligence Labs (USA)",
      "parameters": "Approximately 29.6B including approximately 1.8B perception encoder",
      "context_window": "131,072 tokens",
      "license": "Apache License 2.0",
      "key_features": "Open-weight multimodal agentic model distilled from Muse Spark and optimized by Meta for autonomous workflows on consumer hardware. Muse Glimmer uses a 52-layer dense causal Transformer with a 6,656-dimensional hidden state, SwiGLU feed-forward layers, 32 query heads and 2 key-value heads, gated attention, and a repeating three-local/one-global attention pattern; local layers use a 2,048-token sliding window, while the official configuration supports 131,072 tokens overall. A dedicated approximately 1.8B-parameter, 50-layer ViT-G/14 perception encoder supports interleaved text and image input with text output and up to 4,096 visual tokens per image. Meta trained the model for end-to-end agentic task completion, schema-based tool calling, multi-step reasoning, failure recovery, coding, and compatibility with agent scaffolds, while also supporting more than 100 languages and low, medium, high, and xhigh reasoning-strength controls. Meta reports scores including 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 51.2 on SWE-Bench Pro, 76.0 on SWE-Bench Verified, 77.0 on IFBench, and 94.7 on AIME 2026, positioning it strongly against Gemma4-31B and Qwen3.6-27B in its size class. The release includes BF16 weights, K-Quant-Dynamic and K-Quant-17GB GGUF variants, a frozen perception encoder, and a five-layer DFlash block-diffusion drafter that proposes 16-token blocks for speculative decoding. Meta says the approximately 4-bit quantizations reduce the language model below 20 GB with measured average benchmark degradation of 0.2% for K-Quant-Dynamic and 1.0% for K-Quant-17GB; with DFlash, it reports 233.4 tokens per second and a 3.1x speedup on an RTX 5090. Official artifacts target local and edge deployment through GGUF and ExecuTorch, with integrations announced for llama.cpp, MLX, vLLM, SGLang, Ollama, LM Studio, and Unsloth.",
      "paper_url": "https://research.meta.ai/static/muse-glimmer-methodology",
      "links": [
        "https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model",
        "https://huggingface.co/meta-models/Muse-Glimmer-30B",
        "https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF",
        "https://huggingface.co/meta-models/Muse-Glimmer-30B-assistant",
        "https://huggingface.co/meta-models/Muse-Glimmer-30B-ExecuTorch-PTE",
        "https://research.meta.ai/static/muse-glimmer-methodology",
        "https://developer.meta.com/ai/models/muse-glimmer/",
        "https://www.reddit.com/r/LocalLLaMA/comments/1vkgsum/introducing_muse_glimmer_an_openweight_model/"
      ],
      "page_url": "https://nlp.com.ai/models/muse-glimmer-30b/",
      "record_updated_at": "2026-10-11T08:37:45.748083+00:00"
    },
    {
      "slug": "motif-3",
      "source_file": "2026-08-10-motif-3.json",
      "release_date": "2026-08-10",
      "date": "August 10, 2026",
      "name": "Motif-3",
      "developer": "Motif Technologies (South Korea)",
      "parameters": "~314B total / ~13.2B active per token (sparse MoE)",
      "context_window": "262,144 tokens",
      "license": "MIT License",
      "key_features": "Final MIT-licensed release of Motif Technologies' fully in-house large-scale sparse Mixture-of-Experts language model, following the July Motif-3-Beta checkpoint. Motif-3 is a 53-layer decoder-only architecture with 2 dense layers followed by 51 MoE layers, 384 routed experts with top-8 activation plus 1 shared expert, a 4,096-dimensional hidden state, and approximately 13.2B of its 314B parameters active per token. Its custom architecture centers on Grouped Differential Latent Attention (GDLA), which combines grouped differential attention with the compressed key-value representation of Multi-head Latent Attention and a query-dependent output gate. It also introduces Expert-Specific PolyNorm activations, modified manifold-constrained hyper-connections that mix four parallel residual streams, and a one-layer Multi-Token Prediction head that supports self-speculative decoding. The model was pretrained on approximately 12.5T tokens spanning web documents, STEM, code, mathematics, multilingual and domain-specialized corpora, with added emphasis on Korean, reasoning-intensive, legal, and financial data. Post-training combines general supervised fine-tuning, six reinforcement-learning specialist teachers, a separately trained software-engineering teacher, and Multi-teacher On-Policy Distillation to consolidate reasoning, coding, professional work, calibrated abstention, instruction following, and long-horizon tool use into one model. Motif reports scores including 38.7 on GDPVal v2, 35.3 on tau3-Banking, 51.5 on ITBench, 76.2 on SWE-Bench Verified, 74.9 on Terminal-Bench 2.1, 83.4 on GPQA Diamond, 37.0 on Humanity's Last Exam, 72.3 on AA-LCR, and 78.2 on IFBench. Official artifacts include the post-trained BF16 checkpoint, an unaligned base checkpoint intended for continued training and research, and an NVFP4 quantization validated for serving on two NVIDIA B200 GPUs. Motif reports that NVFP4 remains within approximately 0.15 average points of its Block-FP8 deployment across five accuracy evaluations. The release supports native 256K context, automatic tool selection with Motif-specific reasoning and tool-call parsers, online Block-FP8 quantization, expert-parallel vLLM deployment, and built-in MTP self-speculative decoding.",
      "paper_url": "https://arxiv.org/abs/2608.09119",
      "links": [
        "https://huggingface.co/Motif-Technologies/Motif-3",
        "https://huggingface.co/Motif-Technologies/Motif-3-Base",
        "https://huggingface.co/Motif-Technologies/Motif-3-NVFP4",
        "https://huggingface.co/collections/Motif-Technologies/motif-3",
        "https://arxiv.org/abs/2608.09119",
        "https://motiftech.io/"
      ],
      "page_url": "https://nlp.com.ai/models/motif-3/",
      "record_updated_at": "2026-10-11T08:37:45.745781+00:00"
    },
    {
      "slug": "ling-30-tiny",
      "source_file": "2026-08-10-ling-3.0-tiny.json",
      "release_date": "2026-08-10",
      "date": "August 10, 2026",
      "name": "Ling-3.0-tiny",
      "developer": "inclusionAI (China)",
      "parameters": "7.9B total / 1.3B active (MoE)",
      "context_window": "262,144 tokens with YaRN (131,072 native)",
      "license": "MIT",
      "key_features": "Compact native hybrid-reasoning model from Ant Group's inclusionAI, designed for local, edge, and lightweight agentic deployment with only 1.3B of its 7.9B parameters active per token. Ling-3.0-tiny uses a 3:1 hybrid-linear attention pattern comprising three Kimi Delta Attention layers followed by one Multi-Head Latent Attention layer in each four-layer block. Its sparse Mixture-of-Experts architecture contains 128 routed experts, activates eight routed experts plus one shared expert per token, and supports both fast non-thinking responses and multi-step reasoning through a per-request thinking control. The published deployment recipe extends its native 131,072-token context to 262,144 tokens with YaRN and includes structured tool calling, MTP/NEXTN speculative decoding, prefix caching, and Ling-specific SGLang and vLLM support. inclusionAI reports scores of 25 on Artificial Analysis Intelligence Index v4.1.1, 16 on the Artificial Analysis Agentic Index, 772 Elo on GDPval-AA v2, and 20.80 on tau3-Banking, with Artificial Analysis measuring output above 160 tokens per second. Local validation reports approximately 100-105 tokens per second on an NVIDIA DGX Spark and 86-90 tokens per second on an M4 Pro MacBook, using about 8.34 GiB of peak memory at an 8K context. The release provides official BF16, FP8, and INT4 checkpoints under the MIT license.",
      "paper_url": "",
      "links": [
        "https://x.com/AntLingAGI/status/2086854604621554039",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny-fp8",
        "https://huggingface.co/inclusionAI/Ling-3.0-tiny-int4",
        "https://openrouter.ai/inclusionai/ling-3.0-tiny:free",
        "https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-tiny",
        "https://github.com/inclusionAI/vllm-ling-v3"
      ],
      "page_url": "https://nlp.com.ai/models/ling-30-tiny/",
      "record_updated_at": "2026-10-11T08:37:45.743401+00:00"
    },
    {
      "slug": "mach-1-additive-35b-text-multimodal-mach-1-small",
      "source_file": "2026-08-05-mach-1-additive-35b.json",
      "release_date": "2026-08-05",
      "date": "August 5, 2026",
      "name": "Mach-1 Additive 35B (Text, Multimodal / Mach-1 Small)",
      "developer": "Syzygy Research",
      "parameters": "35B total / approximately 3B active (MoE)",
      "context_window": "262,144 tokens",
      "license": "Apache License 2.0",
      "key_features": "Custom-packed additive compression of Qwen3.6-35B-A3B, released in a text-focused repository and a same-day multimodal variant that adds the base model's BF16 vision tower to the byte-identical compressed language payload. The underlying 40-layer Mixture-of-Experts architecture contains 256 routed experts, selects eight experts per token, and repeats three linear-attention layers followed by one full-attention layer. The published configuration specifies a 2,048-dimensional hidden state, 16 attention heads, two key-value heads, and a 262,144-token context window. Syzygy's manifest reports a 7.5335 GB text payload averaging 1.6978 bits per weight, including 6.2083 GB of packed experts. Expert matrices use integer L1-ball trellis codes with per-wavefront scales and continuous FP16 side streams; the non-expert spine uses a 64-level integer lattice, while the language-model head uses grouped INT5 and the embedding uses INT4. Syzygy describes the representation as enabling add/subtract-only weight matrix multiplications. The release drops the source model's multi-token-prediction head but retains its reasoning-oriented chat template and structured tool-calling format. A bundled standalone NumPy decoder documents and verifies reconstruction of the custom weight representation, although the packed files are not directly loadable as an ordinary Transformers Safetensors checkpoint. Syzygy reports 95.0% mean score retention relative to the BF16 Qwen3.6-35B-A3B teacher across 12 benchmarks in its evaluation harness. Published absolute scores include 89.58 on AIME26, 98.00 on MATH-500, 87.50 on AIME25, 94.44 on MBPP+, 92.68 on HumanEval+, 83.75 on IFEval, 68.97 on BFCL-v3, 71.58 on tau2-bench, and 54.08 on IFBench. The multimodal variant adds the uncompressed 27-layer Qwen vision tower; Syzygy states that its compressed language weights are otherwise byte-identical to the text release.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B",
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B-Multimodal",
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B/blob/main/BOARD.md",
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B/blob/main/MANIFEST.json",
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B/blob/main/RELEASE_NOTES.md",
        "https://huggingface.co/SyzygyResearch/Mach-1-Additive-35B/blob/main/decode.py",
        "https://withsyzygy.com/",
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B"
      ],
      "page_url": "https://nlp.com.ai/models/mach-1-additive-35b-text-multimodal-mach-1-small/",
      "record_updated_at": "2026-10-11T08:37:45.741157+00:00"
    },
    {
      "slug": "maple-preview-maple-preview-flash",
      "source_file": "2026-08-04-maple-preview.json",
      "release_date": "2026-08-04",
      "date": "August 4, 2026",
      "name": "Maple-Preview / Maple-Preview (Flash)",
      "developer": "DeepGrove (USA)",
      "parameters": "20.2B total / 1.49B active (MoE)",
      "context_window": "131,072 tokens",
      "license": "MIT",
      "key_features": "Open-weight preview of DeepGrove's reasoning-focused Maple family, designed from the beginning for fast, memory-efficient inference on consumer devices. Maple-Preview is a 20.2B-total, 1.49B-active ternary-weight Mixture-of-Experts model with 24 layers, 256 routed experts, eight experts selected per token, a 2,048-dimensional hidden state, 16 attention heads, and four key-value heads. Its attention stack repeats three 512-token sliding-window layers followed by one global-attention layer, bounding most KV-cache growth while supporting a 131,072-token context window. The model's weights use ternary values represented as {-alpha, 0, +alpha}; DeepGrove provides the original Safetensors checkpoint, a compressed 5.31 GB 2-bit MLX package, and official TQ1_0 and TQ2_0 GGUF variants under the MIT license. Maple-Preview and Maple-Preview (Flash) share the same underlying transformer weights. The standard mode computes exact logits through the dense output head, while the optional FlashHead artifact clusters the 151,936-token vocabulary into 4,748 clusters and computes exact logits only for the top 512 candidate clusters, with designated special tokens always scored. DeepGrove reports 169 decode tokens per second with the exact head and 218 tokens per second with FlashHead on a 16 GB M4 Mac mini; on an M5 Pro, it reports 359 and 395 tokens per second respectively. Its published reasoning evaluation gives the exact-head model 75.1 on LCBv6, 87.5 on AIME 2026, 78.8 on HMMT 2026, and 73.5 on GPQA-Diamond, averaging 78.7. The FlashHead configuration scores 75.6, 85.8, 80.3, and 69.2 respectively, averaging 77.7. DeepGrove cautions that this preview received minimal agentic post-training and only small-scale general reinforcement learning, so it is focused primarily on raw reasoning and may underperform on agentic tasks ahead of Maple's planned full release.",
      "paper_url": "",
      "links": [
        "https://deepgrove.ai/maple-preview",
        "https://huggingface.co/deepgrove/maple-preview",
        "https://huggingface.co/deepgrove/maple-preview-2bit-mlx",
        "https://huggingface.co/deepgrove/maple-preview-GGUF",
        "https://github.com/deepgrove-ai/mlx-lm-deepgrove",
        "https://github.com/deepgrove-ai/llama.cpp",
        "https://deepgrove.ai/maple-inference"
      ],
      "page_url": "https://nlp.com.ai/models/maple-preview-maple-preview-flash/",
      "record_updated_at": "2026-10-11T08:37:45.738835+00:00"
    },
    {
      "slug": "ling-30-flash",
      "source_file": "2026-08-04-ling-3.0-flash.json",
      "release_date": "2026-08-04",
      "date": "August 4, 2026",
      "name": "Ling-3.0-flash",
      "developer": "inclusionAI (China)",
      "parameters": "124B total / 5.1B active (MoE)",
      "context_window": "262,144 tokens",
      "license": "MIT",
      "key_features": "Next-generation native hybrid-reasoning language model from Ant Group's inclusionAI, designed for efficient production-scale agents and released with only 5.1B of its marketed 124B parameters active per token. Ling-3.0-flash uses a 42-layer hybrid-linear Mixture-of-Experts architecture comprising 35 Kimi Delta Attention layers and 7 gated Multi-head Latent Attention layers in a repeating 5:1 pattern, with two initial dense layers followed by a fine-grained 1/64 sparse MoE. The architecture contains 512 routed experts plus one shared expert, selects eight routed experts per token, and uses a 2,560-dimensional hidden state with 32 attention heads. Its pretraining context schedule progresses from 8K to 32K and finally 256K, while the published configuration supports 262,144 tokens. The model combines thinking and non-thinking behavior in one checkpoint, with thinking enabled by default, and provides native structured tool calling for multi-turn agent workflows. inclusionAI says its agent-focused training used more than 10,000 interactive environments spanning coding, general agent tasks, and deep research, alongside a hierarchical SGLang HiCache and Mooncake caching design intended to reduce redundant computation in long-running interactions. Official reported results include 93.2 on AIME 2026, 87.0 on HMMT February 2026, 72.4 on SWE-Bench Multilingual, 56.6 on SWE-Bench Pro, and 22.7 on Humanity's Last Exam. The release includes official BF16, FP8, FP4, and INT4 repositories under the MIT license, custom BailingMoeV3 model code, MTP/NEXTN speculative decoding support, and deployment instructions for SGLang and inclusionAI's Ling-specific vLLM branch.",
      "paper_url": "",
      "links": [
        "https://x.com/AntLingAGI/status/2084656533489754475",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-fp8",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-fp4",
        "https://huggingface.co/inclusionAI/Ling-3.0-flash-int4",
        "https://developer.ant-ling.com/en/docs/models/ling/",
        "https://github.com/inclusionAI/vllm-ling-v3"
      ],
      "page_url": "https://nlp.com.ai/models/ling-30-flash/",
      "record_updated_at": "2026-10-11T08:37:45.736653+00:00"
    },
    {
      "slug": "lfm25-26b-base-agentic",
      "source_file": "2026-08-04-lfm2.5-2.6b.json",
      "release_date": "2026-08-04",
      "date": "August 4, 2026",
      "name": "LFM2.5-2.6B (Base, Agentic)",
      "developer": "Liquid AI (USA)",
      "parameters": "2.69B",
      "context_window": "131,072 tokens",
      "license": "LFM Open License v1.0",
      "key_features": "Open-weight dense text-model release comprising the pretrained LFM2.5-2.6B-Base checkpoint and LFM2.5-2.6B, an always-thinking post-trained model designed for agentic workloads, tool use, data extraction, retrieval-augmented generation, and long-context workflows on consumer hardware. The model uses Liquid AI's hybrid LFM2 architecture with 30 layers: 22 double-gated short-convolution blocks and eight grouped-query-attention layers. Its configuration specifies a 2,048-dimensional hidden state, 32 attention heads, eight key-value heads, a 128,000-token vocabulary, and a 131,072-token context window. Liquid AI reports a 34-trillion-token training budget, including a mid-training stage that extended context to 128K. Agentic post-training consists of two supervised fine-tuning rounds, per-domain teacher specialization, multi-domain on-policy distillation, and reinforcement learning conducted inside popular agent harnesses so the model encounters their tools, system prompts, and interaction patterns. The model supports 16 languages, ChatML-style conversations, structured function calling, and explicit reasoning through think tags. Published benchmark results include 51.87 on AIME25, 59.41 on LiveCodeBench v6, 59.17 on IFBench, 80.07 on Multi-IF, 56.88 on BFCLv4, 77.83 on ToolSandbox, and 26.89 on BrowseComp+ with OpenClaw. Liquid AI reports decode speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 while using under 2.5 GB of memory, plus nearly 15,000 output tokens per second at high concurrency on a single H100. Official native, GGUF, ONNX, and MLX releases support Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio, and ONNX Runtime. Liquid AI does not recommend the model for agentic coding or knowledge-heavy tasks.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://www.liquid.ai/blog/lfm2-5-2-6b",
        "https://huggingface.co/LiquidAI/LFM2.5-2.6B",
        "https://huggingface.co/LiquidAI/LFM2.5-2.6B-Base",
        "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF",
        "https://huggingface.co/LiquidAI/LFM2.5-2.6B-ONNX",
        "https://huggingface.co/LiquidAI/LFM2.5-2.6B-MLX",
        "https://arxiv.org/abs/2511.23404"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-26b-base-agentic/",
      "record_updated_at": "2026-10-11T08:37:45.734349+00:00"
    },
    {
      "slug": "bigbang-v1",
      "source_file": "2026-08-02-bigbang-v1.json",
      "release_date": "2026-08-02",
      "date": "August 2, 2026",
      "name": "BigBang-V1",
      "developer": "The Endless Frontier",
      "parameters": "35.95B total / approximately 3B active (MoE)",
      "context_window": "262,144 tokens",
      "license": "Apache License 2.0",
      "key_features": "General-purpose multimodal model evolved from Qwen3.6-35B-A3B through post-training for scientific research, reasoning, coding, and tool use. BigBang-V1 was trained on approximately 10,000 high-difficulty examples produced by a self-evolving synthetic-data framework in which generator agents create and solve problems while critic agents assess correctness, difficulty, scalability, and diversity. It uses a 40-layer hybrid architecture that repeats three linear-attention layers followed by one full-attention layer. Its Mixture-of-Experts design contains 256 experts, selects eight experts per token, and activates approximately 3B parameters. The model supports reasoning, structured tool calls, multimodal image and video input, and multi-token prediction. It includes a 27-layer vision encoder and can be deployed in text-only mode without loading the vision encoder. The Endless Frontier reports scores of 76.5 on BrowseComp, 54.2 on SWE-Bench Pro, 50.3 on Humanity's Last Exam, and 46.2 on FrontierScience Research. Official deployment guidance covers Transformers, SGLang, vLLM, and KTransformers.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/endless-frontier/BigBang-v1",
        "https://huggingface.co/endless-frontier/BigBang-v1/blob/main/bigbang_main_figure.pdf",
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B"
      ],
      "page_url": "https://nlp.com.ai/models/bigbang-v1/",
      "record_updated_at": "2026-10-11T08:37:45.731860+00:00"
    },
    {
      "slug": "longcat-flash-lite-sparse",
      "source_file": "2026-08-01-longcat-flash-lite-sparse.json",
      "release_date": "2026-08-01",
      "date": "August 1, 2026",
      "name": "LongCat-Flash-Lite-Sparse",
      "developer": "Meituan LongCat (China)",
      "parameters": "69B total / ~3B active per token (MoE)",
      "context_window": "Up to 1M tokens",
      "license": "MIT",
      "key_features": "Non-thinking open-weight Mixture-of-Experts language model built on LongCat-Flash-Lite, replacing its dense Multi-head Latent Attention with LongCat Sparse Attention and repeating the long-context extension stage on an enriched corpus. LongCat Sparse Attention combines streaming-aware indexing, which reserves fixed sink and local-window selections for more contiguous KV-cache access; cross-layer indexing, which reuses one indexing pass across adjacent layers; and optional hierarchical indexing, which narrows candidates through coarse-to-fine scoring. The model contains 69 billion total parameters while activating approximately 3 billion per token and natively supports context lengths up to one million tokens. Meituan reports substantial gains over the dense LongCat-Flash-Lite predecessor on agentic coding, search, and tool-use evaluations, including 68.2 on SWE-bench Verified without hierarchical indexing versus 54.4 for the dense model, and 95.18 on tau2 Telecom versus 72.8. The release provides approximately 138 GB of BF16 Safetensors across 26 shards, a Transformers chat template, and an SGLang deployment path capable of serving the model on a single H20 141 GB node. Hierarchical indexing can be enabled at inference time without additional training, allowing deployments to trade some benchmark quality for lower indexing overhead.",
      "paper_url": "https://arxiv.org/abs/2608.01662",
      "links": [
        "https://huggingface.co/meituan-longcat/LongCat-Flash-Lite-Sparse",
        "https://arxiv.org/abs/2608.01662",
        "https://github.com/sgl-project/sglang/pull/32918"
      ],
      "page_url": "https://nlp.com.ai/models/longcat-flash-lite-sparse/",
      "record_updated_at": "2026-10-11T08:37:45.729489+00:00"
    },
    {
      "slug": "openpangu-20-pro",
      "source_file": "2026-07-31-openpangu-2.0-pro.json",
      "release_date": "2026-07-31",
      "date": "July 31, 2026",
      "name": "openPangu-2.0-Pro",
      "developer": "Huawei (China)",
      "parameters": "505B total / 18B active (MoE)",
      "context_window": "524,288 tokens",
      "license": "OpenPangu Model License Agreement Version 2.0",
      "key_features": "Ascend-native open-weight Mixture-of-Experts language model designed for long-context agentic reasoning, with unified thinking and non-thinking behavior, structured tool calling, and a native 512K-token context window. Huawei specifies 505B total parameters with approximately 18B active per token, while the Hugging Face repository indexes 541.49B stored tensor parameters across the shipped checkpoint, which includes three multi-token-prediction layers and auxiliary components. The 50-layer architecture begins with three dense layers followed by 47 MoE layers, using 384 routed experts plus one shared expert and selecting eight routed experts per token. Its configuration specifies a 5,120-dimensional hidden state, 64 attention heads, a 1,792-dimensional expert intermediate state, and a 151,552-token vocabulary. openPangu-2.0-Pro builds on Multi-Head Latent Attention with a repeating 1:2 mixture of DeepSeek Sparse Attention and Sliding-Window Attention. Sliding-window layers use a 512-token local window, while sparse global layers use a lightweight indexer to select 2,048 relevant historical tokens. Additional attention mechanisms include 128 learned parametric attention-sink entries per layer and local-context ModAttn convolutions around global attention. The model replaces a single residual stream with four-stream Manifold-Constrained Hyper-Connections, adds sandwich and periodic block normalization for stability, and was pretrained with the Muon optimizer on approximately 34 trillion tokens without reported loss spikes. Three MTP heads predict three additional tokens for self-speculative decoding. Post-training consists of unified supervised fine-tuning, parallel reinforcement-learning specialists for reasoning, general tasks, agents, and coding, followed by multi-expert on-policy distillation to combine those capabilities. Huawei reports thinking-mode scores of 95.4 on AIME 2026, 86.2 on HMMT February 2026, 87.9 on GPQA-Diamond, 27.1 on Humanity's Last Exam, 81.1 on TAU2-Bench, 65.7 on BrowseComp, 85.7 on LiveCodeBench v6, and 68.5 on SWE-bench Verified. Its Ascend-optimized omni-infer deployment stack includes specialized mHC, sparse-attention, communication, cache-management, prefix-caching, tokenization, and quantization optimizations; the technical report claims 9.55 ms time per output token at ultra-low latency and 1,326 generated tokens per second per NPU at 20 ms TPOT with 128K context. The OpenPangu Model License Agreement Version 2.0 prohibits accessing, downloading, installing, running, deploying, integrating, modifying, or otherwise using the model within the European Union and requires specified openPangu attribution when distributing products or services based on the model.",
      "paper_url": "https://huggingface.co/openpangu/openPangu-2.0-Pro/blob/main/openPangu-2.0%20Tech%20Report.pdf",
      "links": [
        "https://www.huawei.com/cn/news/2026/6/open-pangu-flash-model",
        "https://huggingface.co/openpangu/openPangu-2.0-Pro",
        "https://ai.gitcode.com/ascend-tribe/openPangu-2.0-Pro",
        "https://gitcode.com/ascend-tribe/openPangu-2.0-Infer",
        "https://huggingface.co/openpangu/openPangu-2.0-Pro/blob/main/openPangu-2.0%20Tech%20Report.pdf",
        "https://huggingface.co/openpangu/openPangu-2.0-Pro/blob/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/openpangu-20-pro/",
      "record_updated_at": "2026-10-11T08:37:45.727127+00:00"
    },
    {
      "slug": "deepseek-v4-flash-0731",
      "source_file": "2026-07-31-deepseek-v4-flash-0731.json",
      "release_date": "2026-07-31",
      "date": "July 31, 2026",
      "name": "DeepSeek-V4-Flash-0731",
      "developer": "DeepSeek AI (China)",
      "parameters": "284B total / 13B active (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "MIT",
      "key_features": "Official DeepSeek-V4-Flash release that supersedes the April preview while retaining the same 284B-total, 13B-active Mixture-of-Experts architecture and receiving a new round of post-training focused heavily on agentic performance. The published configuration describes a 43-layer DeepSeek-V4 causal architecture with 256 routed experts plus one shared expert, six routed experts selected per token, 64 attention heads, compressed sparse-attention components, YaRN context scaling from an original 65,536-token window to 1,048,576 tokens, FP8 weight quantization, and an attached DSpark speculative-decoding module. The model supports low, high, and max reasoning-effort settings; DeepSeek recommends up to 384K output tokens for high and max effort and provides dedicated OpenAI-compatible message encoding and output parsing rather than a Jinja chat template. DeepSeek reports substantially stronger agent capabilities than both V4-Flash Preview and V4-Pro Preview, including 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 25.2 on Agents' Last Exam, and 25.1 on AutomationBench Public. Its internal results include 68.7 on DSBench-FullStack and 59.6 on DSBench-Hard. DeepSeek positions the release as broadly competitive with leading proprietary agent models despite its smaller active parameter count, with native Responses API support and specific adaptation for Codex-style agent workflows. The official checkpoint bundles the target and DSpark draft weights together and includes deployment guidance for vLLM, SGLang, and DeepSeek's local inference implementation.",
      "paper_url": "https://arxiv.org/abs/2606.19348",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731",
        "https://api-docs.deepseek.com/updates/",
        "https://arxiv.org/abs/2606.19348",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/tree/main/encoding",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/tree/main/inference",
        "https://api-docs.deepseek.com/guides/thinking_mode"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v4-flash-0731/",
      "record_updated_at": "2026-10-11T08:37:45.724870+00:00"
    },
    {
      "slug": "barunlm-35m",
      "source_file": "2026-07-31-barunlm-35m.json",
      "release_date": "2026-07-31",
      "date": "July 31, 2026",
      "name": "BarunLM-35M",
      "developer": "Harshal Singh (India)",
      "parameters": "35,072,768",
      "context_window": "2,048 tokens",
      "license": "Apache License 2.0",
      "key_features": "Compact English-centric decoder-only base language model designed to maximize capability per parameter and released as a native PyTorch completion model rather than an instruction-tuned chat model. BarunLM-35M uses 12 dense Transformer layers with a 448-dimensional hidden state, seven query heads, one shared key-value head, a 1,228-dimensional feed-forward layer, and tied input/output embeddings over a 16,384-token byte-level BPE vocabulary. Its hybrid attention schedule repeats three local-attention layers with a 256-token window followed by one full-attention layer, providing periodic global information exchange while limiting most KV-cache growth. Additional architectural features include grouped-query attention, 50% partial RoPE, query-key normalization, sigmoid-gated attention outputs, activation-bounded SwiGLU, and a learned convex residual selector after every four-layer block that can preserve the block input when its transformation is unhelpful. The model was pretrained on 5,699,985,408 tokens, approximately 162.5 tokens per parameter, from a deduplicated mixture of FineWeb-Edu, Cosmopedia v2, FineMath-4+, DCLM-Baseline, CodeSearchNet Python, and CodeParrot Clean. Its final four-billion-token continuation used Muon with a peak learning rate of 1e-4, 0.1 weight decay, a batch size of 48, and 40,690 optimizer steps on one NVIDIA H200. Harshal Singh reports 41.01% macro accuracy across a fixed nine-task zero-shot LM Evaluation Harness suite comprising ARC-Challenge, ARC-Easy, BoolQ, HellaSwag, LAMBADA OpenAI, OpenBookQA, PIQA, SciQ, and WinoGrande. Under the same decontaminated protocol, BarunLM reportedly exceeds LFM2.5-230M-Base by 1.81 percentage points while using 6.55 times fewer parameters; the published 10,000-resample bootstrap interval for that difference is +0.92 to +2.71 points. Evaluation excludes 1,854 samples detected through a correctness-blind exact 13-token scan of the complete training history. The release includes the checkpoint, tokenizer, architecture implementation, generation script, checksums, configuration, and detailed benchmark evidence under Apache 2.0. Its creator cautions that the model's small capacity, short context, lack of instruction tuning, and English-centric training limit factual recall, multi-step reasoning, instruction following, and high-stakes use.",
      "paper_url": "",
      "links": [
        "https://x.com/HarshalsinghCN/status/2083223503554449745",
        "https://huggingface.co/harrrshall/BarunLM-35M",
        "https://github.com/harrrshall/barunlm-35m",
        "https://harrrshall.github.io/barunlm-35m/blog/",
        "https://huggingface.co/harrrshall/BarunLM-35M/blob/main/barun_config.json",
        "https://huggingface.co/harrrshall/BarunLM-35M/blob/main/benchmark_results.json",
        "https://huggingface.co/harrrshall/BarunLM-35M/blob/main/CITATION.cff"
      ],
      "page_url": "https://nlp.com.ai/models/barunlm-35m/",
      "record_updated_at": "2026-10-11T08:37:45.722433+00:00"
    },
    {
      "slug": "inkling-small-bf16-nvfp4",
      "source_file": "2026-07-30-inkling-small.json",
      "release_date": "2026-07-30",
      "date": "July 30, 2026",
      "name": "Inkling-Small (BF16, NVFP4)",
      "developer": "Thinking Machines Lab, Inc. (United States)",
      "parameters": "276B total / 12B active (MoE)",
      "context_window": "1M tokens",
      "license": "Apache 2.0",
      "key_features": "Efficient open-weight multimodal reasoning model that delivers performance comparable to the 975B-total, 41B-active Inkling flagship at roughly one quarter of its active and total parameter counts. Inkling-Small is a 42-layer decoder-only autoregressive transformer with a sparse Mixture-of-Experts feed-forward backbone: each token is routed to 6 of 256 experts alongside 2 shared experts, while a hybrid of local and global attention layers supports context windows of up to 1M tokens. It uses the same natively multimodal, encoder-free design as Inkling, jointly processing text with image representations produced by a four-layer hierarchical patch encoder and audio represented through discrete dMel spectrogram tokens; accepted inputs are text, images, and audio, with text output. The final model follows an earlier preview checkpoint and incorporates an improved pretraining data mix and training recipe, on-policy distillation from Inkling, and two additional weeks of scaled agentic-coding reinforcement learning. It supports variable thinking effort so deployments can trade reasoning compute and output length against performance. Thinking Machines reports that the final release surpasses Inkling on several reasoning and agentic-coding evaluations while retaining competitive multimodal performance, including 80.2% on SWE-Bench Verified, 55.9% on SWE-Bench Pro Public, 64.7% on Terminal-Bench 2.1, 89.5% on GPQA Diamond, 31.6% on Humanity's Last Exam text-only, 47.8% on Humanity's Last Exam with tools, 95.5% on AIME 2026, 40.1% on ARC-AGI-2, 74.0% on MMMU Pro, 54.9% on Audio MC, and 90.1% on VoiceBench at its published high-effort evaluation setting. The full checkpoint is available in BF16 and an official NVFP4 quantization, with documented local deployment support through SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face Transformers.",
      "paper_url": "",
      "links": [
        "https://thinkingmachines.ai/news/inkling-small/",
        "https://thinkingmachines.ai/model-card/inkling-small/",
        "https://huggingface.co/thinkingmachines/Inkling-Small",
        "https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4"
      ],
      "page_url": "https://nlp.com.ai/models/inkling-small-bf16-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.719910+00:00"
    },
    {
      "slug": "ax-k2-fp8",
      "source_file": "2026-07-29-a-x-k2.json",
      "release_date": "2026-07-29",
      "date": "July 29, 2026",
      "name": "A.X K2 (FP8)",
      "developer": "SK Telecom (South Korea)",
      "parameters": "688B total / 33B active (MoE)",
      "context_window": "262,144 tokens (128K native; extended to 256K via YaRN)",
      "license": "Apache 2.0",
      "key_features": "Flagship open-weight successor to A.X K1, developed as part of the South Korean government's Sovereign AI Foundation Model project and designed for Korean- and English-language reasoning, coding, tool use, knowledge work, and long-context applications. A.X K2 is a text-only, decoder-only Mixture-of-Experts transformer with 688B reported total parameters and 33B active parameters per token. Its 61-layer architecture consists of one dense layer followed by 60 MoE layers, with 256 routed experts, eight selected experts per token, one shared expert, a 7,168-dimensional hidden state, 64 attention heads, and a 163,840-token vocabulary. Hugging Face's checkpoint metadata counts approximately 691.69B stored tensor elements, slightly above SK Telecom's architectural headline of 688B. The model combines Multi-head Latent Attention with SK Telecom's Sparse Gated Attention, whose lightweight indexer selects the top 2,048 relevant positions for each query, and head-specific output gating intended to suppress attention sinks and improve long-context efficiency. Gated Norm applies an input-dependent gate after RMSNorm to suppress activation outliers and stabilize low-precision training and serving. Its Think-Fusion post-training recipe places controllable thinking and non-thinking behavior in one checkpoint through paired supervised examples followed by multi-stage on-policy reinforcement learning; the included chat template also supports structured tool calling and retention of reasoning across turns. Pretraining used approximately 8.2T tokens in a three-stage curriculum comprising 6.4T general-knowledge tokens, 1.4T higher-quality reasoning tokens, and 0.36T long-context tokens, with substantial English, Korean, and code coverage plus smaller Chinese, Japanese, and Spanish components. Post-training included two supervised stages using about 60.6B tokens of paired thinking and non-thinking examples, followed by multi-task and safety reinforcement learning over roughly 100K prompts. The model was trained natively with MXFP8 E4M3 forward and backward computation and released as a block-scaled FP8 checkpoint using 128-by-128 weight blocks with dynamic activation scaling. SK Telecom reports 97.1 on AIME26, 80.5 on KMMLU-Pro, 91.6 on CLIcK, 84.0 on LiveCodeBench v6, 85.6 on GPQA Diamond, 27.8 on Humanity's Last Exam, 75.9 on IFBench, and 98.0 on the telecom portion of tau-squared Bench. It scores a 94.6 overall average on RULER through 256K context, while separate needle-in-a-haystack testing reports perfect retrieval through 512K using more aggressive YaRN scaling.",
      "paper_url": "https://github.com/SKT-AI/A.X-K2/blob/main/A_X_K2_Tech_Report.pdf",
      "links": [
        "https://news.sktelecom.com/en/3204",
        "https://huggingface.co/skt/A.X-K2",
        "https://github.com/SKT-AI/A.X-K2",
        "https://huggingface.co/docs/transformers/main/en/model_doc/axk2"
      ],
      "page_url": "https://nlp.com.ai/models/ax-k2-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.717503+00:00"
    },
    {
      "slug": "neutrino-1-8b-06b-06b-chat",
      "source_file": "2026-07-27-neutrino-1.json",
      "release_date": "2026-07-27",
      "date": "July 27, 2026",
      "name": "Neutrino-1 (8B, 0.6B, 0.6B-Chat)",
      "developer": "Fermion Research",
      "parameters": "8.19B; 596M (0.6B and 0.6B-Chat)",
      "context_window": "40,960 tokens",
      "license": "Apache 2.0 weights; bundled optimized binaries have separate proprietary terms",
      "key_features": "Grouped release of three compact open-weight language models trained for Fermion Research's proprietary ternary-family representation: the general-purpose Neutrino-1 8B chat model, a Neutrino-1 0.6B base and speculative-drafting model, and the conversationally post-trained Neutrino-1 0.6B-Chat. The 8B model is derived from Qwen3-8B through ternary quantization-aware training and staged post-training for knowledge, instruction following, mathematics, tool use, and clean termination. It contains exactly 8,190,735,360 parameters across 36 decoder layers, with a 4,096-dimensional hidden state, 12,288-dimensional gated SwiGLU feed-forward blocks, 32 query heads, eight key-value heads, a 151,936-token vocabulary, and a 40,960-token context window. Its 252 attention and feed-forward projection matrices contain approximately 6.95B ternary weights, while its separate input and output embedding tables remain int8 and its normalization tensors remain full precision. The serving container occupies approximately 3.88 GB and its losslessly compressed transport download is approximately 2.56 GB, compared with roughly 16 GB for an equivalently sized FP16 checkpoint. The two 0.6B variants contain exactly 596,049,920 parameters and retain the same tokenizer and context length in a narrower 28-layer architecture with a 1,024-dimensional hidden state, 3,072-dimensional feed-forward blocks, 16 query heads, eight key-value heads, and tied int8 embeddings. The base 0.6B can generate independently or draft tokens for the 8B without vocabulary remapping; its repository was updated on July 31 with a draft-specialist checkpoint distilled from 8B rollouts, while the original launch checkpoint remains available through revision pinning. Fermion reports that greedy speculative decoding preserves the 8B's output exactly, with a launch certification covering 27,648 generated tokens without divergence. Published launch evaluations for the 8B include 72.1 on five-shot MMLU, 67.8 on generative MMLU-Redux, 77.2 prompt-strict on IFEval, 68.9 macro across thirteen BFCL v3 subsets, and 53.4 on zero-shot GSM8K using flexible extraction. Fermion reports 396 tokens per second for plain decoding and up to 763 tokens per second with the 0.6B draft on an H100, plus local inference through a shared Python loader, MLX pack, public llama.cpp fork, GGUF-compatible custom tensor format, and bundled x86 and Apple-silicon runtimes. The weights are Apache-2.0, but Fermion's optimized prebuilt runtime binaries remain closed-source and carry separate redistribution and reverse-engineering restrictions.",
      "paper_url": "",
      "links": [
        "https://www.fermionresearch.com/research/neutrino-8b/",
        "https://www.fermionresearch.com/research/one-eighth-the-bits/",
        "https://www.fermionresearch.com/research/the-neutrino-engine/",
        "https://huggingface.co/FermionResearch/Neutrino-8B",
        "https://huggingface.co/FermionResearch/Neutrino-0.6B",
        "https://huggingface.co/FermionResearch/Neutrino-0.6B-Chat",
        "https://github.com/fermionresearch/llama.cpp"
      ],
      "page_url": "https://nlp.com.ai/models/neutrino-1-8b-06b-06b-chat/",
      "record_updated_at": "2026-10-11T08:37:45.715153+00:00"
    },
    {
      "slug": "kimi-k3",
      "source_file": "2026-07-27-kimi-k3.json",
      "release_date": "2026-07-27",
      "date": "July 27, 2026",
      "name": "Kimi K3",
      "developer": "Moonshot AI (China)",
      "parameters": "2.8T total, 104B active (MoE)",
      "context_window": "1,048,576 tokens",
      "license": "Kimi K3 License",
      "key_features": "Moonshot AI's open-weight native multimodal agentic flagship for long-horizon coding, reasoning, knowledge work, visual tasks, and tool-driven workflows. Kimi K3 is a 93-layer, 2.8-trillion-parameter Mixture-of-Experts model with 104B activated parameters, making it the first publicly released open-weight model in the 3T parameter class. Its architecture combines 69 Kimi Delta Attention layers with 24 gated Multi-head Latent Attention layers, Attention Residuals for information flow across model depth, and Stable LatentMoE. The MoE contains 896 routed experts with 16 selected per token plus two shared experts, alongside one dense layer, a 7,168-dimensional attention hidden state, 96 attention heads, a 160K vocabulary, and the SiTU-GLU activation function. Moonshot says these changes provide approximately 2.5 times the overall scaling efficiency of Kimi K2. Native multimodal support covers text, images, and video through a 401M-parameter MoonViT-V2 vision encoder, while the 1,048,576-token context window targets enormous repositories, document collections, extended agent trajectories, and multimodal knowledge work. The released checkpoint was quantization-aware trained from the supervised-fine-tuning stage onward with MXFP4 weights and MXFP8 activations and occupies approximately 1.56 TB on Hugging Face. Post-training includes reinforcement learning across general, coding, agentic, and multiple reasoning-effort settings, supported by infrastructure for million-token agentic RL with persistent rollout and sandbox state. Kimi K3 always operates in thinking mode, exposes low, high, and max reasoning-effort levels with max as the default, and requires preserved reasoning history across multi-turn conversations and tool calls. Moonshot's maximum-effort evaluations report 93.5 on GPQA Diamond, 67.5 on DeepSWE, 88.3 on Terminal-Bench 2.1, 77.8 on ProgramBench, 81.2 on FrontierSWE, 42.0 on SWE-Marathon, 91.2 on BrowseComp with context management, 84.2 on MCP-Atlas, 63.3 on OfficeQA Pro, 91.1 on OmniDocBench, and 90.0 on Video-MME with subtitles. Official deployment guidance covers vLLM, SGLang, and TokenSpeed. The custom Kimi K3 License broadly permits use, modification, distribution, fine-tuning, and sale, but commercial model-as-a-service operators exceeding $20 million in aggregate revenue over a consecutive 12-month period must obtain a separate Moonshot agreement; very large commercial products also carry a Kimi K3 display requirement.",
      "paper_url": "https://arxiv.org/abs/2607.24653",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-K3",
        "https://huggingface.co/collections/moonshotai/kimi-k3",
        "https://github.com/MoonshotAI/Kimi-K3",
        "https://www.kimi.com/blog/kimi-k3",
        "https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k3/",
      "record_updated_at": "2026-10-11T08:37:45.712717+00:00"
    },
    {
      "slug": "instella-moe-16b-a3b-pretrain-midtrain-base-sft-dpo-think",
      "source_file": "2026-07-24-instella-moe-16b-a3b.json",
      "release_date": "2026-07-24",
      "date": "July 24, 2026",
      "name": "Instella-MoE-16B-A3B (Pretrain, Midtrain, Base, SFT, DPO, Think)",
      "developer": "AMD (USA)",
      "parameters": "16B total / 2.8B active per token (MoE)",
      "context_window": "Up to 64K tokens (4K before the long-context Base stage)",
      "license": "Research-Only RAIL Model License (academic and research purposes only)",
      "key_features": "Fully disclosed Mixture-of-Experts language-model family trained from scratch by AMD, grouping six downloadable checkpoints spanning Pretrain, Midtrain, long-context Base, supervised fine-tuning, direct preference optimization, and reinforcement-learned Think stages. The 27-layer decoder has a 2,048-dimensional hidden state, 16 attention heads, 64 routed experts, two shared experts, and six routed experts selected per token, giving it 16 billion total parameters with approximately 2.8 billion active. Its architecture adds a learned output gate to Multi-head Latent Attention, called Gated MLA, and uses FarSkip-Collective connectivity to overlap expert-parallel communication with computation; AMD reports a 12.7% pretraining speedup and up to a 39.2% reduction in time to first token under expert-parallel SGLang serving. Pretraining covered 7.1 trillion tokens of openly identified web, mathematics, code, science, and instruction-style data on AMD Instinct MI300X and MI325X accelerators using ROCm and Primus, with a Multi-Token Prediction objective during the Pretrain and Midtrain stages. The Base checkpoint extends context from 4K to 64K using YaRN, document-boundary masking, and Longmino 100B data, after which AMD provides SFT, DPO, and Think checkpoints for instruction following and chain-of-thought reasoning. The final RL process uses the Miles framework, 1,400 steps of instruction-following-specialized training, and Multi-Teacher On-Policy Distillation to incorporate instruction-following gains while preserving math, coding, and general capabilities. AMD reports a 76.7 average for the Base checkpoint across its standard evaluation suite, including 86.5 on WinoGrande and 65.7 on HumanEval+, plus 41.5 on HELMET and 79.4 on RULER at long context. The Think checkpoint reaches a reported 73.22 aggregate score, 83.70 on IFEval, 82.50 on AGIEval, and 73.40 on AIME25. In addition to every major weight checkpoint, AMD released training configurations, data-mixture documentation, intermediate artifacts, tokenizer assets, and inference and training code. The model weights are downloadable but restricted by AMD's Research-Only RAIL license to academic and research purposes, with mandatory responsible-use conditions.",
      "paper_url": "",
      "links": [
        "https://rocm.blogs.amd.com/artificial-intelligence/instella-moe/README.html",
        "https://huggingface.co/collections/amd/instella-moe",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-Pretrain",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-Midtrain",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-Base",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-SFT",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-DPO",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-Think",
        "https://github.com/AMD-AGI/Instella-MoE",
        "https://huggingface.co/amd/Instella-MoE-16B-A3B-Think/blob/main/LICENSE",
        "https://openreview.net/forum?id=ruOpvLzsGV"
      ],
      "page_url": "https://nlp.com.ai/models/instella-moe-16b-a3b-pretrain-midtrain-base-sft-dpo-think/",
      "record_updated_at": "2026-10-11T08:37:45.710063+00:00"
    },
    {
      "slug": "kat-coder-v25-dev-bf16",
      "source_file": "2026-07-23-kat-coder-v2.5-dev.json",
      "release_date": "2026-07-23",
      "date": "July 23, 2026",
      "name": "KAT-Coder-V2.5-Dev (BF16)",
      "developer": "KwaiKAT Team at Kuaishou Technology (China)",
      "parameters": "35B total / 3B active (MoE; 34.66B checkpoint parameters)",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight developer-oriented member of the KAT-Coder-V2.5 family, post-trained specifically for autonomous repository-level software engineering, agentic coding, tool use, and long-horizon debugging workflows. KAT-Coder-V2.5-Dev is built by post-training Qwen3.6-35B-A3B and retains its hybrid Qwen3.5-MoE architecture identifier. The checkpoint contains approximately 34.66B BF16 parameters, marketed as 35B total with 3B activated per token. Its 40-layer language backbone alternates three GatedDeltaNet linear-attention layers with one full-attention layer, yielding 30 linear-attention and 10 full-attention layers. It uses a 2,048-dimensional hidden state, 16 attention heads, two key-value heads, 256 routed experts with eight selected per token, one shared expert, and a 262,144-token native context window. Although the inherited configuration declares image and video components, Kwaipilot explicitly states that the released checkpoint contains only language-model weights and must be served in text-only mode; no vision tower or multimodal capability is included. The model supports both default thinking behavior and a lower-latency non-thinking mode, structured tool calling, and optional preservation of reasoning traces across turns for multi-step agent workflows. Post-training begins with supervised fine-tuning on 127K examples and continues with ten stable reinforcement-learning epochs. The RL pipeline retains the broader KAT-V2.5 framework's token-in/token-out consistency, truncated importance sampling, reliability-hardened sandboxes and verifiers, and hierarchical rewards derived from execution feedback, while adding penalties for excessive parallel tool calls, failed or empty tool calls, and repeated content. Kwaipilot reports that this reduced malformed tool labels from 9.34% to 0.28% and eliminated measured single-turn continuous repetition from 0.34% to zero. Under its standardized in-house evaluation harness, the team reports scores of 69.40 on SWE-bench Verified, 63.00 on SWE-bench Multilingual, 45.96 on SWE-bench Pro, 41.02 averaged across two Terminal-Bench 2.1 harnesses, 93.43 on PinchBench, 44.20 on SciCode, and 46.21 on KAT-Code-Bench. The official artifacts support deployment through Transformers, vLLM, SGLang, and KTransformers, with vLLM requiring its language-model-only option because the vision weights are absent.",
      "paper_url": "https://arxiv.org/abs/2607.05471",
      "links": [
        "https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev",
        "https://github.com/Kwaipilot/KAT-Coder",
        "https://streamlake.com/product/kat-coder"
      ],
      "page_url": "https://nlp.com.ai/models/kat-coder-v25-dev-bf16/",
      "record_updated_at": "2026-10-11T08:37:45.707639+00:00"
    },
    {
      "slug": "solar-open-2-250b-a15b",
      "source_file": "2026-07-22-solar-open-2-250b.json",
      "release_date": "2026-07-22",
      "date": "July 22, 2026",
      "name": "Solar Open 2 (250B-A15B)",
      "developer": "Upstage (South Korea)",
      "parameters": "250B total / 15B active per token (MoE)",
      "context_window": "1M tokens",
      "license": "Upstage Solar License",
      "key_features": "Official open-weight flagship from Upstage positioned as a Korean sovereign foundation model optimized for agentic work, long-context efficiency, and practical office and coding use rather than pure benchmark spectacle. The official model card and technical report describe Solar Open 2 as a 250B-A15B sparse Mixture-of-Experts model with 48 layers, 320 routed experts plus 1 shared expert, and a hybrid attention stack that interleaves one softmax layer with three linear-attention layers repeatedly across the network. Compared with Solar Open 1, the new release enlarges the expert pool, replaces RoPE with a NoPE design, extends context from 128K to 1M, and uses selective weight transfer from only about 2.3% of Solar Open 1’s compatible weights to accelerate training while preserving architectural freedom. Upstage says the linear-attention layers maintain sequence information in fixed-size state rather than growing KV cache linearly, enabling about one quarter of the memory and computation of an all-softmax stack at long context. The report also emphasizes that the model was pretrained on about 12T tokens with a 4:6 real-to-synthetic data mix, more than 80% English, at least 15% each of code and math, and a final 0.9T-token length-expansion phase that injects reasoning and agent data while extending context to 1M. Public benchmark tables position Solar Open 2 as a serious open-weight contender across both English and Korean tasks, including MMLU-Pro, GPQA-Diamond, HLE, LiveCodeBench, ArtifactsBench, APEX-Agents, Ko-GDPval, KMMLU-Pro, CLIcK, and other Korean reasoning and office-agent benchmarks, with claims that it roughly matches or approaches much larger models like DeepSeek-V4-Pro on some Korean agent tasks at a fraction of the size. The release also highlights support for agentic workflows, tool calling, and multilingual English/Korean/Japanese use, with public deployment guidance through Hugging Face, vLLM, and SGLang.",
      "paper_url": "https://huggingface.co/upstage/Solar-Open2-250B/resolve/main/Solar_Open_2_Tech_Report.pdf?download=true",
      "links": [
        "https://huggingface.co/upstage/Solar-Open2-250B",
        "https://huggingface.co/upstage/Solar-Open2-250B/resolve/main/Solar_Open_2_Tech_Report.pdf?download=true",
        "https://www.upstage.ai/blog/en/solar-open-2",
        "https://www.upstage.ai/blog/ko/solar-open-2",
        "https://www.upstage.ai/"
      ],
      "page_url": "https://nlp.com.ai/models/solar-open-2-250b-a15b/",
      "record_updated_at": "2026-10-11T08:37:45.705034+00:00"
    },
    {
      "slug": "nanbeige42-3b",
      "source_file": "2026-07-22-nanbeige4.2-3b.json",
      "release_date": "2026-07-22",
      "date": "July 22, 2026",
      "name": "Nanbeige4.2-3B",
      "developer": "Nanbeige (China)",
      "parameters": "4B total / 3B non-embedding parameters",
      "context_window": "131,072 max new tokens; evaluation and assistant usage emphasize long multi-turn thinking and tool workflows",
      "license": "Apache-2.0",
      "key_features": "Official compact agentic release from Nanbeige, positioned as a 3B-class local personal assistant and code agent that outperforms significantly larger open models on a broad in-house and public benchmark mix. The official model card says Nanbeige4.2-3B is built on Nanbeige4.2-3B-Base and uses a Looped Transformer architecture that reuses transformer layers to increase effective capacity without adding parameters. Compared with Nanbeige4.1-3B, the release shifts emphasis from general reasoning alone toward stronger agentic behavior, with supervised fine-tuning expanded through real-world environment integrations, large-scale environment synthesis, diversified task assets and scaffolds, and trajectory- and turn-level filtering using both test-case validation and rubric-based assessment. The card also says reinforcement learning combines outcome and process rewards to stabilize compact-model training. Public benchmark tables position Nanbeige4.2-3B ahead of larger models such as Qwen3.5-9B and Gemma4-12B on multiple general-agent, code-agent, and local-assistant tasks, including GDPval rubrics, Agent-IF-Oneday, Office-QA-Pro, Pinch-Bench-V2, Claw-Gym, Claw-Eval, MCP-Atlas, SWE-Bench Verified, SWE-Bench Pro, and Terminal-Bench 2.0. The release further highlights OpenClaw-oriented local assistant use, configurable thinking and preserve_thinking controls in the chat template, XML or JSON tool-call formatting, and custom parser support across Nanbeige-maintained SGLang, vLLM, llama.cpp, and Ollama integrations. The modeling code also points forward to Nanbeige4.5, where the team says LoopSplit, depth-attentive mHC, and concatenated n-gram embeddings are already incorporated for a later 2026 release.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Nanbeige/Nanbeige4.2-3B",
        "https://huggingface.co/Nanbeige",
        "https://huggingface.co/Nanbeige/Nanbeige4.2-3B-Base"
      ],
      "page_url": "https://nlp.com.ai/models/nanbeige42-3b/",
      "record_updated_at": "2026-10-11T08:37:45.702614+00:00"
    },
    {
      "slug": "macaron-v1-venti-coding-venti-tall",
      "source_file": "2026-07-21-macaron-v1.json",
      "release_date": "2026-07-21",
      "date": "July 21, 2026",
      "name": "Macaron-V1 (Venti, Coding-Venti, Tall)",
      "developer": "Mind Lab",
      "parameters": "Venti/Coding-Venti: 748B aggregate (744B base + four 1B LoRA specialists), ~40B active; Tall: 50B aggregate (35B-A3B base + four 3.7B LoRA specialists)",
      "context_window": "1M tokens (Venti and Coding-Venti); 262K tokens (Tall)",
      "license": "MIT",
      "key_features": "Official non-preview Macaron-V1 release grouping three open-weight text-generation checkpoints: the flagship Macaron-V1-Venti, the directly usable coding-specialized Macaron-V1-Coding-Venti, and the smaller Macaron-V1-Tall. Venti is described as the first post-trained derivative of GLM-5.2 and combines its sparse Mixture-of-Experts base with four 1B-parameter LoRA specialists, while Tall applies the same system to Qwen3.6-35B-A3B using four rank-64 LoRA specialists of approximately 3.7B parameters each. The Mixture-of-LoRA architecture divides work among L0 Chat, L1 Agent, L2 Coding, and L3 GenUI specialists; L0 routes each new request, ongoing reasoning and tool interactions remain with the selected specialist, and concise summaries communicate completed work between their separate contexts. Coding-Venti merges the coding specialist into the GLM-5.2 base so it can be loaded as a conventional coding checkpoint without the routed multi-LoRA harness. Venti and Coding-Venti support up to a 1M-token context, while Tall's released configuration supports 262K tokens. Post-training used Mind Lab's MinT infrastructure and MindForge recursive self-improvement loop, with LongStraw providing an experimental execution path for reinforcement learning beyond two million tokens under fixed GPU budgets. Mind Lab evaluated V1 across ChatBench, LivingBench, VitaBench, PinchBench, SWE-bench Verified, DeepSWE, SWE Atlas QnA, TerminalBench 2.1, and UI4A-Bench, claiming that Venti matches or exceeds the listed proprietary frontier baselines across most tracks while remaining closest to, rather than ahead of, the frontier on coding. The release includes BF16 base weights and adapters, chat templates, an MIT-licensed OpenAI-compatible routing proxy, and deployment profiles for native multi-LoRA serving through vLLM or SGLang.",
      "paper_url": "",
      "links": [
        "https://alpha.macaron.im/mindlab/research/introducing-macaron-v1",
        "https://huggingface.co/collections/mindlab-research/macaron-v1",
        "https://huggingface.co/mindlab-research/Macaron-V1-Venti",
        "https://huggingface.co/mindlab-research/Macaron-V1-Coding-Venti",
        "https://huggingface.co/mindlab-research/Macaron-V1-Tall",
        "https://github.com/MindLab-Research/Mixture-of-LoRA-Harness",
        "https://arxiv.org/abs/2607.14952"
      ],
      "page_url": "https://nlp.com.ai/models/macaron-v1-venti-coding-venti-tall/",
      "record_updated_at": "2026-10-11T08:37:45.700213+00:00"
    },
    {
      "slug": "laguna-s-21-bf16-fp8-nvfp4-int4-gguf-dflash",
      "source_file": "2026-07-21-laguna-s-2.1.json",
      "release_date": "2026-07-21",
      "date": "July 21, 2026",
      "name": "Laguna S 2.1 (BF16, FP8, NVFP4, INT4, GGUF, DFlash)",
      "developer": "Poolside (United States)",
      "parameters": "118B total / 8B active per token (MoE)",
      "context_window": "1M tokens",
      "license": "OpenMDW-1.1",
      "key_features": "Official open-weight coding and long-horizon agent release from Poolside, positioned as the company’s most capable model to date and as a compact frontier-coding model that punches above much larger systems. The main model card and release blog describe Laguna S 2.1 as a 118B total-parameter Mixture-of-Experts model with 8B activated parameters per token, built with the same Laguna architecture family as XS 2.1: a token-choice router with softplus gating over 256 routed experts plus one shared expert, grouped-query attention, and interleaved full and sliding-window attention. Poolside says S 2.1 went from the start of training to launch in under nine weeks, supports both thinking and no-thinking modes up to 1M context, and adds broad deployment formats including BF16, FP8, NVFP4, INT4, GGUF, and a 1B DFlash draft model for speculative decoding. Release materials emphasize aggressive long-horizon coding performance at relatively low active compute, with Poolside reporting 70.2 on Terminal-Bench 2.1, 78.5 on SWE-Bench Multilingual, 59.4 on SWE-Bench Pro Public, 40.4 on DeepSWE, 46.2 on SWE Atlas Codebase QnA, and 49.7 on Toolathlon Verified, while openly publishing full evaluation trajectories for the final checkpoint. Poolside’s blog frames the release around stronger post-training, improved reasoning use, and rapid iteration in its internal Model Factory stack, while the public serving docs highlight reasoning controls, native tool-calling support, and compatibility with Transformers, vLLM, SGLang, TRT-LLM, llama.cpp, and Ollama workflows.",
      "paper_url": "",
      "links": [
        "https://poolside.ai/blog/introducing-laguna-s-2-1",
        "https://huggingface.co/poolside/Laguna-S-2.1",
        "https://huggingface.co/poolside/Laguna-S-2.1-FP8",
        "https://huggingface.co/poolside/Laguna-S-2.1-NVFP4",
        "https://huggingface.co/poolside/Laguna-S-2.1-GGUF",
        "https://huggingface.co/collections/poolside/laguna-s-21",
        "https://trajectories.poolside.ai/"
      ],
      "page_url": "https://nlp.com.ai/models/laguna-s-21-bf16-fp8-nvfp4-int4-gguf-dflash/",
      "record_updated_at": "2026-10-11T08:37:45.697800+00:00"
    },
    {
      "slug": "antares-350m-1b",
      "source_file": "2026-07-21-antares.json",
      "release_date": "2026-07-21",
      "date": "July 21, 2026",
      "name": "Antares (350M, 1B)",
      "developer": "Cisco Foundation AI / Cisco Systems, Inc. (United States)",
      "parameters": "350M and 1B dense",
      "context_window": "32K tokens",
      "license": "Apache 2.0",
      "key_features": "Official small-model open-weight release family from Cisco Foundation AI focused narrowly on agentic vulnerability localization rather than general chat or broad reasoning benchmarks. The public model cards describe Antares-350M and Antares-1B as decoder-only Granite 4.0 descendants trained to operate inside a terminal loop, where the model receives only a generic CWE category description and must explore a code repository with shell commands such as grep, find, and cat before submitting likely vulnerable files. Both models are trained with a two-stage pipeline of supervised fine-tuning followed by Group Relative Policy Optimization over full multi-turn agent trajectories, with rewards tied to file-level localization quality, valid terminal behavior, and exploration strategy. Cisco positions the family as compact on-premises security models for CI/CD, advisory-driven triage, and local security workflows, explicitly recommending sandboxed deployment with human oversight rather than use as a general-purpose assistant. The public benchmark tables on the VLoc Bench 500-task evaluation show the family punching above its size in this narrow domain, with Antares-1B (GRPO) reaching a File F1 of 0.209 and Antares-350M (GRPO) reaching 0.135, outperforming several much larger open and closed models on the same protocol, including Gemma-4-31B, GPT-5 Mini, Qwen3.5-122B-A10B, and Gemini 2.5 Flash. Those same tables also reference Antares-3B SFT and GRPO variants that outperform the released pair, but I could not verify a public weight release for the 3B checkpoint. The 350M card further notes that the smallest model can act as a speculative draft model for larger Antares variants, while both cards emphasize local deployment, a 15-command terminal budget, and English-only static training data with an April 10, 2025 cutoff.",
      "paper_url": "https://cisco-foundation-ai.github.io/antares/",
      "links": [
        "https://huggingface.co/fdtn-ai/antares-350m",
        "https://huggingface.co/fdtn-ai/antares-1b",
        "https://huggingface.co/collections/fdtn-ai/antares",
        "https://cisco-foundation-ai.github.io/antares/",
        "https://x.com/Cisco/status/2079552761591484892"
      ],
      "page_url": "https://nlp.com.ai/models/antares-350m-1b/",
      "record_updated_at": "2026-10-11T08:37:45.695222+00:00"
    },
    {
      "slug": "motif-3-beta",
      "source_file": "2026-07-20-motif-3-beta.json",
      "release_date": "2026-07-20",
      "date": "July 20, 2026",
      "name": "Motif-3-Beta",
      "developer": "Motif Technologies (South Korea)",
      "parameters": "~314B total / ~13B active per token (sparse MoE)",
      "context_window": "262,144 tokens",
      "license": "Custom non-commercial research license",
      "key_features": "Official preview checkpoint of Motif-3 from South Korea's Motif Technologies, positioned as a fully in-house large-scale sparse MoE model rather than a repackaged derivative of an existing open architecture. The public Hugging Face model card describes Motif-3-Beta as an intermediate checkpoint ahead of the final Motif-3 release, with about 314B total parameters, roughly 13B active parameters per token, 384 routed experts with top-8 activation plus 1 shared expert, 53 layers, and a native 256K context window. Motif says the architecture introduces several custom components, including Grouped Differential Latent Attention (GDLA), Grouped PolyNorm activations applied per expert, and a modified mHC module. The release is framed as a multilingual general-purpose long-context model and the model card highlights an Artificial Analysis Intelligence Index score of 44, placing it in a serious open-model capability band despite being explicitly labeled as a beta rather than the final checkpoint. Public usage guidance covers Transformers, vLLM, SGLang, and Docker-based serving, while the repo notes that a dedicated vLLM serving guide is still forthcoming and community reports already indicate that the shipped custom modeling code remains rough around the edges. The current release therefore reads as an unusually large public beta artifact from Motif's in-house Korean sovereign-model effort: openly downloadable, architecturally distinctive, but still unfinished enough that the team warns the final checkpoint will follow soon.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Motif-Technologies/Motif-3-Beta",
        "https://huggingface.co/Motif-Technologies/Motif-3-Beta/tree/main",
        "https://motiftech.io/",
        "https://motiftech.io/en/company/"
      ],
      "page_url": "https://nlp.com.ai/models/motif-3-beta/",
      "record_updated_at": "2026-10-11T08:37:45.692623+00:00"
    },
    {
      "slug": "btl-3",
      "source_file": "2026-07-20-btl-3.json",
      "release_date": "2026-07-20",
      "date": "July 20, 2026",
      "name": "BTL-3",
      "developer": "Bad Theory Labs (Nigeria)",
      "parameters": "27B (dense)",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0 model artifacts; MIT native runtime",
      "key_features": "Open-weight agentic coding and structured tool-use family derived from the Qwen3.6-27B hybrid-attention causal language model. The maximum-quality BTL-3 release is the frozen RL-0013 checkpoint distributed as a 933,974,032-byte rank-32, alpha-64 PEFT LoRA adapter for a pinned Qwen3.6-27B base revision, with a 65,536-token maximum reinforcement-learning sequence length and a 32,768-token launch benchmark context. It is trained for repository work, coding, debugging, structured single and parallel function calls, multi-turn execution, recovery after failed actions or verifier feedback, and abstaining when no tool call is appropriate. Bad Theory Labs reports 95.12% HumanEval pass@1, 88.1% on its completed 193-case LiveCodeBench v6 run, 88.5% BFCL v4 AST across 1,240 cases, and 26.35% on BigCodeBench-Hard Instruct; its BFCL results include 95.5% for multiple calls, 87.0% for parallel calls, and 91.2% irrelevance. The coding benchmark results used thinking mode, but the current supported release default is non-thinking because the RL-0013 reasoning policy can become repetitive or fail to terminate. The grouped release also includes BTL-3 Compact, a complete 64-layer standalone text model packaged in a single 8.39 GB GGUF at under 2.5 effective artifact bits per parameter. Compact combines packed AVQ2 and UniSVQ decoder tensors, affine INT4 tensors, selected higher-precision behavior-sensitive islands, packed vocabulary matrices, a rank-32 output correction, and a compact behavior adapter without loading or reconstructing the BF16 base. Its July 24 behavior refresh reports 98.4% conditional retention on an internal 63-case teacher-correct tool-contract gate. The included native runtime directly executes the packed representation on CUDA and Metal and provides an OpenAI-compatible server plus custom LM Studio and Ollama CLI integrations; stock GGUF engines do not directly decode its AVQ2 representation.",
      "paper_url": "https://github.com/Badtheorylabs/BTL-3/blob/main/papers/behavior-before-perplexity/output/pdf/btl-3-behavior-before-perplexity.pdf",
      "links": [
        "https://huggingface.co/badtheorylabs/BTL-3",
        "https://huggingface.co/badtheorylabs/BTL-3-Compact",
        "https://huggingface.co/Qwen/Qwen3.6-27B",
        "https://github.com/Badtheorylabs/BTL-3",
        "https://www.badtheorylabs.com/"
      ],
      "page_url": "https://nlp.com.ai/models/btl-3/",
      "record_updated_at": "2026-10-11T08:37:45.689857+00:00"
    },
    {
      "slug": "intern-s2-preview-397b-bf16-fp8",
      "source_file": "2026-07-17-intern-s2-preview-397b.json",
      "release_date": "2026-07-17",
      "date": "July 17, 2026",
      "name": "Intern-S2-Preview-397B (BF16, FP8)",
      "developer": "Shanghai AI Laboratory / InternLM (China)",
      "parameters": "397B-to-403B class multimodal model (Hugging Face lists approximately 403B parameters)",
      "context_window": "256K tokens text / 64K tokens multimodal evaluation",
      "license": "Apache-2.0",
      "key_features": "Official large-scale successor in the Intern-S2 scientific multimodal family from Shanghai AI Laboratory and the InternLM team, positioned as the project’s most capable open-weight model for scientific intelligence and long-horizon agents. Intern-S2-Preview-397B scales the earlier recipe through a vision-language pre-training paradigm based on raw rendered scientific pages, large-scale multi-task reinforcement learning across more than 20 scientific domains, and long-horizon agent reinforcement learning in sandboxed interactive environments. The full technical report documents scientific multimodal pre-training followed by supervised fine-tuning, scalable multi-task RL, black- and white-box agentic RL, and on-policy distillation, with training-system techniques including partial rollouts with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly. The model learns directly from rendered scientific literature while also using layout-aware interleaved PDF data and large-scale scientific-image retrieval, jointly modeling text, equations, tables, figures, page structure, and other visual evidence. Its upgraded time-series encoder supports approximately 300,000 time steps and is reported to run five to six times faster at maximum length while using around 20% of the previous implementation’s GPU memory; a separate forecasting branch adds direct numerical time-series generation. The separately released Intern-MemDec-4B provides a modular biological-memory extension without modifying the frozen 397B backbone: both branches process the same context, and a lightweight token-level router dynamically fuses their next-token distributions. InternLM reports that this attachment raises the average Biology-Instructions score from 56.92 to 60.32 while broadly preserving general and multimodal capability. Release materials also emphasize native tool calling, thinking and non-thinking modes, scientific generation, and integration with OpenClaw, Hermes, Claude Code, Transformers, vLLM, SGLang, and LMDeploy.",
      "paper_url": "https://arxiv.org/abs/2608.13505",
      "links": [
        "https://huggingface.co/internlm/Intern-S2-Preview-397B",
        "https://huggingface.co/internlm/Intern-S2-Preview-397B-FP8",
        "https://huggingface.co/internlm/Intern-MemDec-4B",
        "https://huggingface.co/collections/internlm/intern-s2",
        "https://github.com/InternLM/Intern-S1",
        "https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview-397B",
        "https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview-397B-FP8",
        "https://arxiv.org/abs/2607.25614"
      ],
      "page_url": "https://nlp.com.ai/models/intern-s2-preview-397b-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.687211+00:00"
    },
    {
      "slug": "llada22-flash",
      "source_file": "2026-07-16-llada2.2-flash.json",
      "release_date": "2026-07-16",
      "date": "July 16, 2026",
      "name": "LLaDA2.2-flash",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "103B total (100B non-embedding, MoE)",
      "context_window": "128K tokens",
      "license": "Apache 2.0",
      "key_features": "Agent-oriented Mixture-of-Experts diffusion language model designed for long-context tool use, multi-turn interaction, and error correction. LLaDA2.2 extends the series from fixed-length token substitution to Levenshtein Editing with KEEP, SUBSTITUTE, DELETE, and INSERT operations, allowing parallel diffusion decoding to remove redundant spans, insert missing positions, and change sequence length. Its L-EBPO reinforcement-learning method combines longest-common-subsequence-derived editing labels with environmental rewards to train editing and recovery behavior during long-horizon agent trajectories. The 32-layer architecture uses 256 routed experts with eight selected per token plus one shared expert, 32 attention heads, a 4,096-dimensional hidden state, RoPE, a 157,184-token vocabulary, and 32-token diffusion blocks. Progressive long-context continual pretraining extends the model to 128K, while Block Routing regulates expert activation at the diffusion-block level to control MoE inference costs. Official evaluations report 80.33 on tau2-Bench, 81.66 on PinchBench, 64.22 on Claw-Eval, 46.21 on MCP-Atlas, 49.28 on SWE-bench Verified, and 60.78 on BFCL-V4. The technical report reports an average 1.64x BF16 throughput advantage over Ling-2.6-flash across 11 tested workloads, including 703.82 tokens per second on BFCL-V4; FP8 quantization provided a further 18.6% average throughput increase. The release supplies approximately 206 GB of BF16 safetensors and custom Transformers implementation code; the model card notes that dedicated SGLang deployment support is still forthcoming.",
      "paper_url": "https://github.com/inclusionAI/LLaDA2.X/blob/main/LLaDA2_2_tech_report.pdf",
      "links": [
        "https://huggingface.co/inclusionAI/LLaDA2.2-flash",
        "https://modelscope.cn/models/inclusionAI/LLaDA2.2-flash",
        "https://github.com/inclusionAI/LLaDA2.X"
      ],
      "page_url": "https://nlp.com.ai/models/llada22-flash/",
      "record_updated_at": "2026-10-11T08:37:45.684735+00:00"
    },
    {
      "slug": "inkling-bf16-nvfp4",
      "source_file": "2026-07-15-inkling.json",
      "release_date": "2026-07-15",
      "date": "July 15, 2026",
      "name": "Inkling (BF16, NVFP4)",
      "developer": "Thinking Machines Lab, Inc. (United States)",
      "parameters": "975B total / 41B active (MoE)",
      "context_window": "1M tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight flagship release from Thinking Machines Lab positioned as a broad multimodal base model for customization, fine-tuning, and agentic product workflows rather than a narrowly benchmark-optimized frontier model. The official announcement and model card describe Inkling as a 66-layer decoder-only multimodal autoregressive transformer with a sparse Mixture-of-Experts feed-forward backbone, routing each token to 6 of 256 experts plus 2 shared experts, with hybrid local/global attention and native joint processing of text, image, audio, and video representations in a shared hidden space. Thinking Machines says the model has 975B total parameters with 41B active parameters, supports up to 1M tokens of context, and was pretrained from scratch on 45T tokens spanning text, images, audio, and video. The release announcement also introduces Inkling-Small as a lighter-weight 12B-active preview trained with a similar recipe, although this entry covers the publicly verifiable main Inkling release. Public materials emphasize controllable thinking effort, strong agentic coding and tool-use behavior across varied harnesses, broad multimodal capability, and compatibility with Tinker for fine-tuning and self-customization workflows, including a demo where Inkling fine-tunes and reloads its own weights. The public benchmark tables position Inkling as a balanced generalist open-weight model with competitive results across reasoning, coding, agentic, factuality, chat, vision, audio, and safety evaluations, including HLE, AIME 2026, GPQA Diamond, SWE-Bench Verified, SWE-Bench Pro Public, Terminal Bench 2.1, MCP Atlas, BrowseComp, Global-MMLU-Lite, MMMU Pro, CharXiv RQ, Audio MC, and VoiceBench. The model card states that the model supports BF16, MXFP8, and NVFP4 numerics, while the public Hugging Face release exposes BF16 and NVFP4 checkpoint formats for local deployment through SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face tooling.",
      "paper_url": "",
      "links": [
        "https://thinkingmachines.ai/news/introducing-inkling/",
        "https://thinkingmachines.ai/model-card/inkling/",
        "https://huggingface.co/thinkingmachines/Inkling"
      ],
      "page_url": "https://nlp.com.ai/models/inkling-bf16-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.682017+00:00"
    },
    {
      "slug": "bonsai-27b-1-bit-ternary-awq-unpacked",
      "source_file": "2026-07-14-bonsai-27b.json",
      "release_date": "2026-07-14",
      "date": "July 14, 2026",
      "name": "Bonsai 27B (1-bit, ternary, AWQ, unpacked)",
      "developer": "Prism ML, Inc.",
      "parameters": "27.3B language parameters + ~0.46B vision tower",
      "context_window": "262K tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Bonsai 27B release family from Prism ML, positioned around extreme intelligence density and on-device deployment rather than conventional full-precision scaling. The public Prism ML materials describe Bonsai 27B as the newest and largest generation in the Bonsai line and the family’s first 27B-class vision-language models, derived from Qwen3.6-27B with the architecture kept unchanged: a hybrid-attention causal model with roughly 75% linear and 25% full attention, SwiGLU MLPs, RoPE, RMSNorm, a 64-block backbone, and an optional ~0.46B vision tower. The release ships multiple public operating points, including native 1-bit Bonsai, ternary Bonsai, AWQ 4-bit, MLX, GGUF, and unpacked safetensors variants. Prism ML says the 1-bit Bonsai 27B model uses binary g128 weights with FP16 group-wise scaling, compressing the deployed language model to about 3.9 GB, while the ternary quality-oriented variant deploys at about 7.2 GB and retains more of the full-precision model’s behavior. Public release materials emphasize 262K-token context, near-lossless 4-bit KV-cache quantization, native OpenAI-style tool calling, MCP support in the demo stack, multimodal image understanding for the 27B models, per-request reasoning effort control, and optional DSpark speculative decoding drafters. The release is framed as the first 27B-class model family able to run on a phone in its native 1-bit format, with published claims of roughly 11 tok/s on an iPhone 17 Pro Max and about 44 tok/s on an M5 Pro laptop, while the ternary family is positioned as the higher-quality default operating point. The public whitepaper and model cards present the family as an efficiency-first, long-context, agent-friendly alternative for local and edge deployment through MLX, llama.cpp, Transformers, vLLM, and SGLang-compatible workflows.",
      "paper_url": "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf",
      "links": [
        "https://prismml.com/",
        "https://github.com/PrismML-Eng/Bonsai-demo",
        "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf",
        "https://huggingface.co/collections/prism-ml/bonsai-27b",
        "https://huggingface.co/prism-ml/Bonsai-27B-unpacked",
        "https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit",
        "https://huggingface.co/prism-ml/Bonsai-27B-AWQ-4bit",
        "https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf"
      ],
      "page_url": "https://nlp.com.ai/models/bonsai-27b-1-bit-ternary-awq-unpacked/",
      "record_updated_at": "2026-10-11T08:37:45.679435+00:00"
    },
    {
      "slug": "agents-a1-4b-base-fp8-gguf",
      "source_file": "2026-07-14-agents-a1-4b.json",
      "release_date": "2026-07-14",
      "date": "July 14, 2026",
      "name": "Agents-A1-4B (base, FP8, GGUF)",
      "developer": "InternScience",
      "parameters": "4B dense",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Official small-model release in the Agents-A1 family from InternScience, positioned as a long-horizon agentic model that brings the series' search, engineering, scientific reasoning, instruction-following, and tool-use behavior into a locally deployable 4B dense checkpoint. The public Agents-A1 paper focuses on the flagship 35B-A3B model and frames the family around scaling agent horizon rather than raw parameter count, using long-horizon knowledge-action trajectories with actions, observations, and verifier outcomes plus multi-teacher domain-routed on-policy distillation across heterogeneous agent domains. The 4B release, announced separately on July 14, 2026, is presented in the model cards as a compact dense variant that still performs strongly on long-horizon search, engineering and research tasks, instruction following, and scientific agentic benchmarks. Official benchmark tables claim clear gains over similarly sized models such as Qwen3.5-4B on BrowseComp, XBench-DS-2510, GAIA, FrontierScience-Research, and IFEval, with some results approaching or surpassing larger MoE models like Nex-N2-mini and Qwen3.6 on selected agentic tasks. The public release also emphasizes native function calling and tool integration, support for SGLang and vLLM with OpenAI-compatible serving, and multimodal capability in the FP8 variant through image-text usage examples. InternScience positions the model as a practical small open agent for local assistants and long-horizon workflows, while tying it back to the broader Agents-A1 training recipe that aims to reach trillion-parameter-level agent performance by scaling trajectories and heterogeneous abilities rather than scaling parameters alone.",
      "paper_url": "https://arxiv.org/abs/2606.30616",
      "links": [
        "https://internscience.github.io/Agents-A1/",
        "https://huggingface.co/InternScience/Agents-A1-4B",
        "https://huggingface.co/InternScience/Agents-A1-4B-FP8",
        "https://huggingface.co/InternScience/Agents-A1-4B-F16-GGUF",
        "https://huggingface.co/collections/InternScience/agents-a1",
        "https://arxiv.org/abs/2606.30616"
      ],
      "page_url": "https://nlp.com.ai/models/agents-a1-4b-base-fp8-gguf/",
      "record_updated_at": "2026-10-11T08:37:45.676696+00:00"
    },
    {
      "slug": "soofi-s-30b-a3b-base-instruct-preview-isar-preview-rhine-preview",
      "source_file": "2026-07-10-soofi-s.json",
      "release_date": "2026-07-10",
      "date": "July 10, 2026",
      "name": "Soofi S 30B-A3B (Base, Instruct Preview, Isar Preview, Rhine Preview)",
      "developer": "Soofi Consortium (Germany)",
      "parameters": "30B total / ~3B active (hybrid MoE; Hugging Face cards also cite ~3.5B active)",
      "context_window": "Unknown",
      "license": "Custom / Other (paper says highly permissive open-access terms; Hugging Face cards still show placeholder license metadata)",
      "key_features": "Official first-generation Soofi release family from a German research consortium, positioned as a sovereign European open-source alternative to US and Chinese foundation models for industrial deployment, code generation, and agentic enterprise use. The public paper and model cards describe Soofi S as a custom hybrid Mamba-2 and Transformer Mixture-of-Experts model trained from scratch in German sovereign infrastructure, with 30B total parameters, roughly 3B active parameters per token, 128 routing experts, 23 Mamba-2/MoE layers plus 6 attention layers, and near-constant inference-cache growth that is framed as a throughput advantage over dense long-context models. The paper says the model was pretrained on about 27 trillion tokens with deliberate German up-weighting and built end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich. Public release materials position the base checkpoint as a research and post-training foundation, while the companion instruction and reasoning variants add assistant behavior, native tool-calling support, and explicit thinking traces for the two reasoning previews, Isar and Rhine. The Soofi paper claims the base model matches dense 14B to 27B models on aggregate English and German benchmarks, achieves the best code aggregates in both languages among 17 open base models, and outperforms other European sovereign baselines including models with far more active parameters. The current Hugging Face release is still explicitly labeled as preview or internal checkpoint material whose weights and metadata may change, and the model cards currently use a custom 'Other' license placeholder while the paper describes the intended release terms as highly permissive open-access terms with weights, intermediate checkpoints, hyperparameters, and training and evaluation code.",
      "paper_url": "https://arxiv.org/abs/2607.09424",
      "links": [
        "https://www.soofi.info/soofi-s/",
        "https://huggingface.co/Soofi-Project/Soofi-S-Base",
        "https://huggingface.co/Soofi-Project/Soofi-S-Instruct-Preview",
        "https://huggingface.co/Soofi-Project/Soofi-S-Isar-Preview",
        "https://huggingface.co/Soofi-Project/Soofi-S-Rhine-Preview",
        "https://huggingface.co/collections/Soofi-Project/soofi-s-beta-models",
        "https://arxiv.org/abs/2607.09424"
      ],
      "page_url": "https://nlp.com.ai/models/soofi-s-30b-a3b-base-instruct-preview-isar-preview-rhine-preview/",
      "record_updated_at": "2026-10-11T08:37:45.673860+00:00"
    },
    {
      "slug": "nemotron-labs-3-puzzle-75b-a9b-bf16-fp8-nvfp4",
      "source_file": "2026-07-06-nemotron-labs-3-puzzle-75b-a9b.json",
      "release_date": "2026-07-06",
      "date": "July 6, 2026",
      "name": "Nemotron-Labs-3-Puzzle-75B-A9B (BF16, FP8, NVFP4)",
      "developer": "NVIDIA (United States)",
      "parameters": "75.3B total / 9.3B active (hybrid MoE)",
      "context_window": "1M tokens",
      "license": "OpenMDW-1.1",
      "key_features": "Official open-weight deployment-optimized Nemotron release from NVIDIA that compresses Nemotron-3-Super-120B-A12B into a smaller hybrid Mamba-Transformer LatentMoE model tuned for interactive serving, reasoning-heavy generation, and long-context agentic workloads. The official model card and paper describe a hybrid architecture with interleaved Mamba, MoE, and attention layers plus Multi-Token Prediction, reduced from the parent model’s 120.7B total / 12.8B active parameters to 75.3B total / 9.3B active through the Iterative Puzzle compression pipeline. NVIDIA says the post-training recipe combines heterogeneous MoE channel pruning, heterogeneous active-expert reduction, Mamba SSM state pruning, long-context knowledge distillation recovery, reinforcement learning focused on software-engineering and agentic skills, and deployment-oriented FP8 and NVFP4 quantization. Release materials claim about 2x higher server throughput than Nemotron-3-Super on a single 8xB200 node at matched user-throughput constraints and an increase from 1 to 8 sustainable concurrent 1M-token requests on a single H100 GPU, while preserving strong performance across reasoning, coding, multilingual, long-context, and agentic benchmarks including MMLU-Pro, AIME25, GPQA, LiveCodeBench, SWE-Bench, Terminal Bench hard subset, TauBench V2, RULER, and WMT24++. NVIDIA positions the model as commercially usable and optimized for collaborative agents, chat systems, RAG applications, and other high-volume production workloads, with published support for Transformers and vLLM on Hopper and Blackwell hardware.",
      "paper_url": "https://arxiv.org/abs/2607.04371",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-FP8",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4",
        "https://arxiv.org/abs/2607.04371",
        "https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.1/LICENSE.OpenMDW-1.1"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-labs-3-puzzle-75b-a9b-bf16-fp8-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.670569+00:00"
    },
    {
      "slug": "hy3-bf16-fp8",
      "source_file": "2026-07-06-hy3.json",
      "release_date": "2026-07-06",
      "date": "July 6, 2026",
      "name": "Hy3 (BF16 + FP8)",
      "developer": "Tencent Hy Team (China)",
      "parameters": "295B total / 21B active (MoE) + 3.8B MTP layer",
      "context_window": "256K tokens",
      "license": "Apache 2.0",
      "key_features": "Official post-preview Hy3 release from Tencent that follows Hy3-preview with scaled-up post-training, higher-quality data, stronger reinforcement learning, and public BF16 plus FP8 open-weight checkpoints. The official model card describes Hy3 as a 295B-parameter Mixture-of-Experts model with 21B active parameters, 3.8B MTP parameters, 192 experts with top-8 activation, 80 main layers, and 256K context. Tencent positions Hy3 as outperforming similar-size models and rivaling much larger flagship open models while improving practical utility in coding, office work, financial modeling, frontend design, game development, and broader productivity tasks. Release materials emphasize stronger agent capabilities, more reliable tool calling and output formats, lower hallucination and commonsense error rates, better long-context and multi-turn intent tracking, and more stable performance across different agent scaffoldings such as CodeBuddy, Cline, and KiloCode. Tencent also highlights a blind evaluation with 270 experts using work-derived tasks, where Hy3 reportedly scored 2.67 out of 4 versus 2.51 for GLM-5.1, plus companion deployment recipes for vLLM and SGLang with MTP-enabled speculative decoding and a released FP8 quantized checkpoint for more practical serving.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/tencent/Hy3",
        "https://huggingface.co/tencent/Hy3-FP8",
        "https://github.com/Tencent-Hunyuan/Hy3",
        "https://modelscope.cn/models/Tencent-Hunyuan/Hy3",
        "https://modelscope.cn/models/Tencent-Hunyuan/Hy3-FP8",
        "https://hunyuan.tencent.com/research/hy3"
      ],
      "page_url": "https://nlp.com.ai/models/hy3-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.667586+00:00"
    },
    {
      "slug": "longcat-20",
      "source_file": "2026-07-05-longcat-2.0.json",
      "release_date": "2026-07-05",
      "date": "July 5, 2026",
      "name": "LongCat-2.0",
      "developer": "Meituan LongCat (China)",
      "parameters": "1.6T total / ~48B active per token (MoE)",
      "context_window": "1M tokens",
      "license": "MIT",
      "key_features": "Official open-weight LongCat flagship release positioned for coding and agentic workflows, with public Hugging Face weights, a published chat template, and deployment guidance for both GPU and NPU environments. The official model card describes LongCat-2.0 as a 1.6 trillion-parameter Mixture-of-Experts model with about 48 billion activated parameters per token, trained on more than 35 trillion tokens and hundreds of billions of 1M-context tokens on AI ASIC superpods rather than Nvidia GPU infrastructure. Release materials emphasize LongCat Sparse Attention, which introduces streaming-aware indexing, cross-layer indexing, and hierarchical indexing to reduce long-context indexing overhead, plus a 135B-parameter N-gram Embedding module inherited from LongCat-Flash-Lite to improve parameter utilization outside the MoE path. Meituan positions the model as deeply compatible with Claude Code, OpenClaw, Hermes, OpenCode, and related developer harnesses, targeting repository-level edits, automated task execution, search, and long-horizon coding work. The official benchmark tables highlight in-house or reported results across Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, FORTE, BrowseComp, RWSearch, IFEval, Writing Bench, IMO-AnswerBench, and GPQA-diamond, presenting LongCat-2.0 as competitive with leading proprietary agentic and coding models while remaining locally deployable with published SGLang support.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/meituan-longcat/LongCat-2.0",
        "https://github.com/meituan-longcat/LongCat-2.0",
        "https://longcat.chat/blog/longcat-2.0/",
        "https://longcat.chat/platform/docs/ChangeLog.html",
        "https://www.modelscope.cn/models/meituan-longcat/LongCat-2.0"
      ],
      "page_url": "https://nlp.com.ai/models/longcat-20/",
      "record_updated_at": "2026-10-11T08:37:45.665137+00:00"
    },
    {
      "slug": "laguna-xs-21-bf16-fp8-nvfp4-int4-gguf",
      "source_file": "2026-07-02-laguna-xs-2.1.json",
      "release_date": "2026-07-02",
      "date": "July 2, 2026",
      "name": "Laguna XS 2.1 (BF16, FP8, NVFP4, INT4, GGUF)",
      "developer": "Poolside (United States)",
      "parameters": "33B total / 3B active per token (MoE)",
      "context_window": "256K tokens",
      "license": "OpenMDW-1.1",
      "key_features": "Official small open-weight coding and long-horizon agent release from Poolside, positioned as a local-machine companion model that upgrades Laguna XS.2 with stronger multilingual software engineering and terminal-style behavior. The Hugging Face card and release blog describe Laguna XS 2.1 as a 33B total-parameter Mixture-of-Experts model with 3B activated parameters per token, using the same token-choice routing setup as XS.2 but a revised architecture built for lower KV-cache pressure and faster inference: grouped-query attention, 40 layers, and sliding-window attention with per-head gating in 30 of those layers. Poolside frames XS 2.1 as the efficient member of the July 2026 Laguna refresh, released alongside Laguna S 2.1, with broad public formats including BF16, FP8, NVFP4, INT4, and GGUF for local deployment. Public benchmark tables emphasize improved coding-agent performance over XS.2 and competitive results for its size, with Poolside reporting 35.0 on Terminal-Bench 2.1, 73.6 on SWE-Bench Multilingual, 42.4 on SWE-Bench Pro Public, 28.5 on DeepSWE, 38.5 on SWE Atlas Codebase QnA, and 32.7 on Toolathlon Verified. The release blog highlights a +5.4 point jump on SWE-Bench Multilingual relative to XS.2, while Poolside’s docs position the model for Poolside Agent CLI workflows and OpenAI-compatible integrations. Overall, the release is pitched as a compact, agent-oriented open model that preserves Laguna’s long-horizon coding identity while making deployment more practical on smaller hardware.",
      "paper_url": "",
      "links": [
        "https://poolside.ai/blog/introducing-laguna-xs-2-1",
        "https://huggingface.co/poolside/Laguna-XS-2.1",
        "https://huggingface.co/poolside/Laguna-XS-2.1-FP8",
        "https://huggingface.co/poolside/Laguna-XS-2.1-GGUF",
        "https://huggingface.co/collections/poolside/laguna-xs2",
        "https://docs.poolside.ai/release-notes/laguna-m1"
      ],
      "page_url": "https://nlp.com.ai/models/laguna-xs-21-bf16-fp8-nvfp4-int4-gguf/",
      "record_updated_at": "2026-10-11T08:37:45.662710+00:00"
    },
    {
      "slug": "ornith-10",
      "source_file": "2026-06-25-ornith-1.0.json",
      "release_date": "2026-06-25",
      "date": "June 25, 2026",
      "name": "Ornith-1.0",
      "developer": "DeepReinforce",
      "parameters": "9B dense, 35B MoE, 397B MoE publicly confirmed; official release page also describes a 31B dense variant",
      "context_window": "262,144 tokens",
      "license": "MIT",
      "key_features": "Official open-source Ornith-1.0 family release focused on agentic coding, built on top of pretrained Gemma 4 and Qwen 3.5 backbones and positioned around a self-improving reinforcement-learning framework that learns both solution rollouts and the scaffolds that guide them. DeepReinforce describes the core training idea as jointly optimizing task solutions and task-specific harnesses rather than relying on fixed human-written agent scaffolds, with the goal of discovering better search trajectories for software-engineering tasks. Public release materials present the family as spanning compact edge-deployable 9B models through frontier-scale 397B MoE variants, while the currently visible Hugging Face release set clearly includes 9B, 35B, and 397B checkpoints plus deployment-oriented FP8 and GGUF variants. The 35B model card describes reasoning-mode behavior with separate reasoning-content parsing, OpenAI-compatible tool calling, and deployment support across Transformers, vLLM, SGLang, llama.cpp, Ollama, and related local-serving stacks. Release benchmarks emphasize strong performance on agentic coding evaluations including Terminal-Bench 2.1, SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual, NL2Repo, ClawEval, and SWE Atlas, with the flagship 397B release positioned against top open models and even compared directly with Claude Opus 4.7 on Terminal-Bench 2.1 and SWE-Bench Verified.",
      "paper_url": "",
      "links": [
        "https://deep-reinforce.com/ornith_1_0.html",
        "https://huggingface.co/deepreinforce-ai",
        "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF"
      ],
      "page_url": "https://nlp.com.ai/models/ornith-10/",
      "record_updated_at": "2026-10-11T08:37:45.660344+00:00"
    },
    {
      "slug": "nemotron-labs-twotower-30b-a3b-base-bf16",
      "source_file": "2026-06-25-nemotron-labs-twotower.json",
      "release_date": "2026-06-25",
      "date": "June 25, 2026",
      "name": "Nemotron-Labs-TwoTower-30B-A3B-Base-BF16",
      "developer": "NVIDIA (United States)",
      "parameters": "~63.2B total shipped parameters across two towers (about 30B AR/context tower + 30B diffusion/denoiser tower); ~3B active per tower",
      "context_window": "128K tokens",
      "license": "NVIDIA Nemotron Open Model License",
      "key_features": "Official open-weight diffusion language model release from NVIDIA that turns the open Nemotron-3-Nano-30B-A3B backbone into a two-tower block-wise autoregressive diffusion generator. The architecture splits generation into a frozen autoregressive context tower and a trainable diffusion denoiser tower, with both towers copied from the same 52-layer hybrid Mamba-2, attention, and MoE backbone. NVIDIA says the denoiser tower adds bidirectional in-block attention, layer-aligned cross-attention to the context tower, context-seeded Mamba state initialization, and timestep conditioning, while keeping the AR tower frozen. The model card says the backbone was pretrained on about 25T tokens, the denoiser adaptation was trained on about 2.1T tokens, and the released checkpoint ships both towers together in BF16. Official benchmarks position the model as preserving 98.7% of the aggregate quality of its autoregressive Nemotron-3-Nano-30B-A3B baseline while delivering 2.42x wall-clock generation throughput at the default operating point, with published comparisons across MMLU, ARC-Challenge, HumanEval, MBPP, GSM8K, MATH-500, and multilingual tasks. NVIDIA also highlights commercial usability, support for Transformers, vLLM, SGLang, and Docker workflows, and a decoding design that generates token blocks iteratively in parallel rather than strictly one token at a time.",
      "paper_url": "https://arxiv.org/abs/2606.26493",
      "links": [
        "https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16",
        "https://huggingface.co/collections/nvidia/nemotron-labs-twotower",
        "https://arxiv.org/abs/2606.26493",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-labs-twotower-30b-a3b-base-bf16/",
      "record_updated_at": "2026-10-11T08:37:45.658001+00:00"
    },
    {
      "slug": "lfm25-230m",
      "source_file": "2026-06-25-lfm2.5-230m.json",
      "release_date": "2026-06-25",
      "date": "June 25, 2026",
      "name": "LFM2.5-230M",
      "developer": "Liquid AI (USA)",
      "parameters": "230M (LFM2 hybrid architecture; 14 layers: 8 double-gated LIV conv + 6 GQA)",
      "context_window": "32,768 tokens",
      "license": "LFM Open License v1.0",
      "key_features": "Open-weight 230M instruction-tuned LFM2.5 release positioned by Liquid AI as its smallest model yet for edge and on-device agentic workflows. Official release materials describe LFM2.5-230M as a text-only hybrid LFM2 model distilled from LFM2.5-350M and then post-trained with a lightweight three-stage recipe: supervised fine-tuning, direct preference optimization, and multi-domain reinforcement learning. The Hugging Face model card lists 19T pretraining tokens, a 32,768-token context window, 65,536 vocabulary size, and support for ten languages including English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, and Spanish. Liquid emphasizes very fast low-footprint inference across constrained hardware, reporting 213 tok/s decode on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5, with day-one support across Transformers, vLLM, SGLang, llama.cpp, MLX, and ONNX companion formats. Benchmark tables in the release compare the model against LFM2.5-350M, prior LFM2-350M, Granite 4.0 350M variants, Qwen3.5-0.8B Instruct, and Gemma 3 1B IT, with especially strong positioning on tool use, function calling, and data extraction tasks such as BFCLv3, BFCLv4, and CaseReportBench. Release guidance recommends it for lightweight on-device agentic pipelines and structured extraction rather than reasoning-heavy math, code generation, or creative writing.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2.5-230M",
        "https://huggingface.co/LiquidAI/LFM2.5-230M-Base",
        "https://www.liquid.ai/blog/lfm2-5-230m",
        "https://arxiv.org/abs/2511.23404",
        "https://docs.liquid.ai"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-230m/",
      "record_updated_at": "2026-10-11T08:37:45.655643+00:00"
    },
    {
      "slug": "qwen-agentworld-35b-a3b",
      "source_file": "2026-06-24-qwen-agentworld-35b-a3b.json",
      "release_date": "2026-06-24",
      "date": "June 24, 2026",
      "name": "Qwen-AgentWorld-35B-A3B",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "35B total / 3B active (MoE)",
      "context_window": "262,144 tokens",
      "license": "Apache-2.0",
      "key_features": "Official open-weight language world model release built on Qwen3.5-35B-A3B-Base and trained specifically for agentic environment simulation rather than general assistant behavior. Qwen describes it as the first language world model to cover seven agent interaction domains within a single model: MCP tool calling, Search, Terminal, SWE, Android, Web, and OS. The Hugging Face model card says the training pipeline runs from continual pretraining through supervised fine-tuning to RL, using world modeling as the objective from the CPT stage onward, and the accompanying paper reports more than 10 million environment-interaction trajectories across seven real-world domains plus a new AgentWorldBench evaluation benchmark. Release materials position the model both as a controllable environment simulator for agent training and as a foundation model whose world-model warmup improves downstream agent performance.",
      "paper_url": "https://arxiv.org/abs/2606.24597",
      "links": [
        "https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B",
        "https://huggingface.co/collections/Qwen/qwen-agentworld",
        "https://qwen.ai/blog?id=qwen-agentworld",
        "https://github.com/QwenLM/Qwen-AgentWorld",
        "https://arxiv.org/abs/2606.24597"
      ],
      "page_url": "https://nlp.com.ai/models/qwen-agentworld-35b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.653184+00:00"
    },
    {
      "slug": "glm-52-bf16-fp8",
      "source_file": "2026-06-16-glm-5.2.json",
      "release_date": "2026-06-16",
      "date": "June 16, 2026",
      "name": "GLM-5.2 (BF16, FP8)",
      "developer": "Z.AI / Zhipu AI (China)",
      "parameters": "~753B total / 40B active (MoE; official repo table also labels the family 744B-A40B)",
      "context_window": "1M",
      "license": "MIT",
      "key_features": "Official open-weight GLM-5.2 flagship release positioned for long-horizon coding and agentic engineering tasks, with both standard and FP8 checkpoints published. Z.AI describes GLM-5.2 as the first GLM release to deliver a 'solid' 1M-token context for practical long-horizon work rather than only nominal long-context support, alongside up to 128K output tokens, multiple reasoning-effort levels, structured output, function calling, MCP integration, and context caching. The official GitHub README says the model introduces IndexShare, which reuses the same sparse-attention indexer across every four sparse layers to reduce per-token FLOPs by 2.9x at 1M context, and also improves the MTP speculative-decoding layer with up to 20 percent longer acceptance length. Official benchmark tables claim GLM-5.2 is the strongest open-source coding model in its comparison set, improving over GLM-5.1 from 62.0 to 81.0 on Terminal-Bench 2.1 and from 58.4 to 62.1 on SWE-bench Pro, while also ranking highly on FrontierSWE, PostTrainBench, SWE-Marathon, MCP-Atlas, and several reasoning benchmarks. Source materials disagree slightly on scale, with Hugging Face listing 753B parameters while the official GitHub download table labels the release family 744B-A40B; both present it as a frontier-scale sparse MoE GLM-5-series checkpoint. The Hugging Face card also emphasizes a pure MIT open-source license with no regional restrictions and local deployment support across SGLang, vLLM, xLLM, Transformers, and KTransformers.",
      "paper_url": "https://arxiv.org/abs/2607.07508",
      "links": [
        "https://huggingface.co/zai-org/GLM-5.2",
        "https://huggingface.co/zai-org/GLM-5.2-FP8",
        "https://z.ai/blog/glm-5.2",
        "https://docs.z.ai/guides/llm/glm-5.2",
        "https://github.com/zai-org/GLM-5",
        "https://github.com/THUDM/slime",
        "https://arxiv.org/abs/2607.07508",
        "https://arxiv.org/abs/2602.15763"
      ],
      "page_url": "https://nlp.com.ai/models/glm-52-bf16-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.650467+00:00"
    },
    {
      "slug": "rio-35-open-397b",
      "source_file": "2026-06-12-rio-3.5-open-397b.json",
      "release_date": "2026-06-12",
      "date": "June 12, 2026",
      "name": "Rio 3.5 Open 397B",
      "developer": "IplanRIO / Prefeitura do Rio de Janeiro (Brazil)",
      "parameters": "397B total / 17B active (MoE-derived release)",
      "context_window": "Unknown",
      "license": "MIT",
      "key_features": "Official open-weight frontier-class release from IplanRIO, the municipal IT company of Rio de Janeiro's city government, published on Hugging Face as a multimodal, agentic general-purpose model. The current official model card presents Rio 3.5 Open 397B as a derivative of Qwen 3.5 397B and says it delivers strong results across agentic coding, mathematics, STEM, multilingual, and multimodal benchmarks while using MIT licensing for redistribution. The model card also describes integration of SwiReasoning, a training-free inference framework based on Shi et al. (2025) that dynamically switches between explicit chain-of-thought and latent-space reasoning using entropy-based confidence signals, with the release framed as explicitly trained to benefit from latent reasoning efficiency. A June 14, 2026 official README correction materially changed the provenance description: it states that the model is built via a merge of Nex-N2-Pro and Qwen3.5-397B-A17B, followed by On-Policy Distillation from a stronger model, and further says an incorrect earlier upload had exposed the base merged model instead of the final distilled release. The Hugging Face page tags the model as image-text-to-text and provides deployment guidance across Transformers, vLLM, SGLang, and Docker Model Runner. Because the currently accessible official page text does not clearly publish a standalone architecture/config section beyond the upstream Qwen lineage, this entry records the officially disclosed derivative provenance rather than inferring unsupported low-level architecture details.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B",
        "https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/commit/a778c1ec4e21180ee55c3ea016a348e549e75f09",
        "https://huggingface.co/Qwen/Qwen3.5-397B-A17B",
        "https://iplanrio.prefeitura.rio"
      ],
      "page_url": "https://nlp.com.ai/models/rio-35-open-397b/",
      "record_updated_at": "2026-10-11T08:37:45.648024+00:00"
    },
    {
      "slug": "minimax-m3",
      "source_file": "2026-06-12-minimax-m3.json",
      "release_date": "2026-06-12",
      "date": "June 12, 2026",
      "name": "MiniMax-M3",
      "developer": "MiniMax (China)",
      "parameters": "~428B total / ~23B active (MoE)",
      "context_window": "1M tokens",
      "license": "minimax-community",
      "key_features": "Official open-weight native multimodal MiniMax flagship release positioned for long-horizon agentic work across coding, cowork, text, image, and video understanding. The Hugging Face model card describes MiniMax-M3 as a million-context multimodal Mixture-of-Experts model with about 428B total parameters and about 23B activated parameters, supporting both thinking and non-thinking modes depending on whether the task prioritizes deeper reasoning or lower-latency completion. The accompanying MiniMax Sparse Attention paper presents M3 as a production model built around blockwise sparse attention co-designed for efficient million-token inference, with group-specific Top-k key-value block selection and a GPU execution path that turns sparsity into practical speedups. Release materials emphasize native multimodal training from the first step rather than late fusion, strong coding and cowork capability on long-horizon agentic benchmarks, and large long-context efficiency gains versus MiniMax M2, including reported 9x prefill and 15x decode speedups at 1M context on the public model card while the technical report cites 14.2x prefill and 7.6x decoding wall-clock speedups on H800 for the MSA kernel versus dense GQA baselines. The card also highlights deployment support across Transformers, vLLM, and SGLang, with recommended inference settings and downloadable open weights for local serving.",
      "paper_url": "https://arxiv.org/abs/2606.13392",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M3",
        "https://arxiv.org/abs/2606.13392"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m3/",
      "record_updated_at": "2026-10-11T08:37:45.645704+00:00"
    },
    {
      "slug": "kimi-k27-code",
      "source_file": "2026-06-12-kimi-k2.7-code.json",
      "release_date": "2026-06-12",
      "date": "June 12, 2026",
      "name": "Kimi-K2.7-Code",
      "developer": "Moonshot AI (China)",
      "parameters": "1T total / 32B active",
      "context_window": "256K",
      "license": "Modified MIT License",
      "key_features": "Official open-weight coding-focused agent model in the Kimi K2 line, positioned by Moonshot AI as a stronger long-horizon software-engineering successor to Kimi K2.6. The model card describes Kimi K2.7 Code as a Mixture-of-Experts architecture with 1T total parameters and 32B activated parameters, 61 layers, 384 experts with 8 selected per token plus 1 shared expert, 64 attention heads, MLA attention, 160K vocabulary, and a 400M-parameter MoonViT vision encoder. Release materials emphasize end-to-end coding and agentic workflows rather than plain code completion, with claimed improvements on real-world long-horizon software-engineering tasks and about 30 percent lower thinking-token usage than Kimi K2.6. The published benchmark table highlights gains over K2.6 on Kimi Code Bench v2, Program Bench, MLS-Bench Lite, Kimi Claw 24/7 Bench, MCP-Atlas, and MCPMark-Verified, while still trailing the strongest frontier closed models on some coding evaluations. Moonshot also says the model forces thinking mode and preserve_thinking by default, retains full reasoning content across turns for coding-agent scenarios, supports interleaved thinking with multi-step tool calling, and accepts image and video inputs in addition to text. The card further notes native INT4 quantization inherited from Kimi-K2-Thinking and deployment support through vLLM, SGLang, KTransformers, and Moonshot's OpenAI/Anthropic-compatible API.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-K2.7-Code"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k27-code/",
      "record_updated_at": "2026-10-11T08:37:45.643309+00:00"
    },
    {
      "slug": "diffusiongemma-26b-a4b-it",
      "source_file": "2026-06-10-diffusiongemma-26b-a4b-it.json",
      "release_date": "2026-06-10",
      "date": "June 10, 2026",
      "name": "DiffusionGemma-26B-A4B-it",
      "developer": "Google DeepMind (USA)",
      "parameters": "25.2B total / 3.8B active (MoE, marketed as A4B)",
      "context_window": "Up to 256K tokens",
      "license": "Apache-2.0",
      "key_features": "Official open-weight Gemma variant that replaces standard token-by-token decoding with discrete text diffusion for exceptionally fast multimodal generation. Google DeepMind created DiffusionGemma by fine-tuning the Gemma 4 Mixture-of-Experts model with 25.2B total and 3.8B active parameters, retaining its text and image understanding, thinking mode, and long-context capabilities. The model uses block-autoregressive multi-canvas sampling: prompt context is processed into a KV cache, after which a decoder with bidirectional attention iteratively denoises 256-token generation canvases in parallel before appending them to the output. Its compute-efficient two-stage training pipeline used fewer than 10% of the starting autoregressive model's training-token budget. The first stage applies supervised fine-tuning to teach bidirectional denoising, while the second combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. The technical report says DiffusionGemma generates approximately 20 tokens per forward pass and averages roughly 1,500 output tokens per second on a single NVIDIA H100, establishing a new capability-versus-speed Pareto frontier against autoregressive models using state-of-the-art speculative decoding. Despite its diffusion fine-tuning, the model can still perform conventional autoregressive generation with only minor reported performance degradation, offering a path toward hybrid diffusion-autoregressive decoding. The instruction-tuned checkpoint also supports native system prompts, optional thinking triggered with the `<|think|>` token, text-to-text and image-to-text tasks, and contexts up to 256K tokens.",
      "paper_url": "https://arxiv.org/abs/2608.00146",
      "links": [
        "https://ai.google.dev/gemma/docs/diffusiongemma",
        "https://ai.google.dev/gemma/docs/diffusiongemma/model_card",
        "https://ai.google.dev/gemma/docs/diffusiongemma/explained",
        "https://ai.google.dev/gemma/docs/diffusiongemma/inference-diffusiongemma-with-hf",
        "https://huggingface.co/google/diffusiongemma-26B-A4B-it",
        "https://arxiv.org/abs/2608.00146"
      ],
      "page_url": "https://nlp.com.ai/models/diffusiongemma-26b-a4b-it/",
      "record_updated_at": "2026-10-11T08:37:45.640852+00:00"
    },
    {
      "slug": "north-mini-code",
      "source_file": "2026-06-09-north-mini-code.json",
      "release_date": "2026-06-09",
      "date": "June 9, 2026",
      "name": "North Mini Code",
      "developer": "Cohere / Cohere Labs (Canada)",
      "parameters": "30B total / 3B active (MoE)",
      "context_window": "256K",
      "license": "Apache 2.0",
      "key_features": "Official open-weight coding-focused release positioned by Cohere as its first model for developers and the first member of a North family of code agent models. Primary release materials describe North Mini Code as a decoder-only sparse Mixture-of-Experts model optimized for code generation, agentic software engineering, and terminal tasks, with 30B total parameters and 3B active parameters. The Hugging Face model card says the architecture uses interleaved sliding-window RoPE attention and global attention in a 3:1 ratio, 128 experts with 8 activated per token, SwiGLU feed-forward experts, and a router that applies sigmoid before top-k selection, plus a single dense layer before the sparse stack. Cohere states the model was post-trained with a two-stage cascaded supervised fine-tuning pipeline followed by reinforcement learning with verifiable rewards focused on agentic coding. Release docs publish a 256K context window with up to 64K output tokens and position the model for repo-level code changes, terminal-based agents, local deployment, and general code-generation tasks. The launch post emphasizes competitive benchmark performance for its size on software-engineering and terminal-agent evaluations, including comparisons against similar open models, and claims up to 2.8x higher output throughput than Devstral Small 2 in Cohere's internal testing under matched hardware and concurrency settings. Distribution is provided through Hugging Face weights, Cohere API, Model Vault, and OpenCode compatibility guidance.",
      "paper_url": "",
      "links": [
        "https://cohere.com/blog/north-mini-code",
        "https://huggingface.co/CohereLabs/North-Mini-Code-1.0",
        "https://docs.cohere.com/docs/north-mini-code-1.0"
      ],
      "page_url": "https://nlp.com.ai/models/north-mini-code/",
      "record_updated_at": "2026-10-11T08:37:45.638434+00:00"
    },
    {
      "slug": "quasar-preview",
      "source_file": "2026-06-08-quasar-preview.json",
      "release_date": "2026-06-08",
      "date": "June 8, 2026",
      "name": "Quasar-Preview",
      "developer": "SILX AI",
      "parameters": "~18B total / ~2B active (MoE)",
      "context_window": "5,000,000 tokens (experimental config)",
      "license": "MIT",
      "key_features": "Official open-weight early preview checkpoint in SILX AI's Quasar Foundation Model series, released as the first public architecture-scale Quasar model for research, decentralized distillation, and Bittensor SN24 training rather than as a finished production release. The model card describes Quasar-Preview as a Quasar Long Hybrid Transformer with roughly 18B total parameters and an approximately 2B active sparse Mixture-of-Experts path, built around hybrid Quasar, Raven, and GLA branches with GQA compatibility attention, Safe NoPE / DrOPE-style long-context staging, and a Loop Transformer scaffold that is present but configured conservatively in this checkpoint. Published technical details list 20 layers, 256 experts with 8 experts active per token, 16 attention heads, 4 KV heads, 157,184 vocabulary size, 64 Raven slots with top-32 routing, and an exposed maximum context configuration of 5,000,000 tokens. Release notes emphasize that the checkpoint is still immature: trained on more than 1T but less than 1.5T tokens overall, with less than 1B tokens of long-context extension training so far, loop execution set to a single pass by default, and Engram memory code included but not active by default. The card also provides an early benchmark snapshot rather than final claims, including reported scores of 68.4 on MMLU, 33.2 on MMLU-Pro, 25.6 on GPQA, 63.0 on ARC Challenge, 80.1 on ARC Easy, 81.9 on PIQA, 74.0 on HellaSwag, 47.0 on OpenBookQA, and 71.4 on MATH-500.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/silx-ai/Quasar-Preview",
        "https://silxinc.com/",
        "https://huggingface.co/silx-ai"
      ],
      "page_url": "https://nlp.com.ai/models/quasar-preview/",
      "record_updated_at": "2026-10-11T08:37:45.636001+00:00"
    },
    {
      "slug": "nex-n2-pro-mini",
      "source_file": "2026-06-08-nex-n2.json",
      "release_date": "2026-06-08",
      "date": "June 8, 2026",
      "name": "Nex-N2 (Pro, mini)",
      "developer": "Nex AGI (China)",
      "parameters": "397B / A17B (Pro); 35B / A3B (mini)",
      "context_window": "256K tokens",
      "license": "Apache-2.0",
      "key_features": "Official open-weight Nex-N2 family release covering Nex-N2-Pro and Nex-N2-mini, positioned as agent models for real-world productivity with a unified Agentic Thinking framework. Nex AGI says the family combines reasoning, tool use, and environment execution into one closed-loop workflow spanning requirement understanding, planning, code implementation, environmental feedback, evaluation, debugging, and iteration, with two explicit design ideas: Adaptive Thinking, where the model decides when and how deeply to reason, and Coherent Thinking, where the same reasoning structure carries across search, coding, terminal tasks, and mixed tool-use workloads. The public release notes say both variants are post-trained on the Qwen3.5 series, with Nex-N2-Pro built on Qwen3.5-397B-A17B and Nex-N2-mini built on Qwen3.5-35B-A3B-Base, targeting different latency and quality tradeoffs. Official benchmark tables emphasize strong agentic and coding results, including reported scores for Nex-N2-Pro of 83.7 on BrowseComp, 1585 on GDPval, 51.9 on Toolathlon, 58.8 on SWE-Bench Pro, 75.3 on Terminal-Bench 2.1, 80.8 on SWE-Bench Verified, and 90.7 on GPQA Diamond, with Nex-N2-mini positioned as the cheaper faster variant while still posting competitive numbers. Release materials also highlight open deployment support through Transformers, vLLM, and a customized SGLang fork, plus explicit reasoning-parser and function-calling guidance for agent harnesses.",
      "paper_url": "",
      "links": [
        "https://nex.sii.edu.cn/",
        "https://huggingface.co/collections/nex-agi/nex-n2",
        "https://huggingface.co/nex-agi/Nex-N2-Pro",
        "https://huggingface.co/nex-agi/Nex-N2-mini"
      ],
      "page_url": "https://nlp.com.ai/models/nex-n2-pro-mini/",
      "record_updated_at": "2026-10-11T08:37:45.633682+00:00"
    },
    {
      "slug": "macaron-v1-preview-749b",
      "source_file": "2026-06-08-macaron-v1-preview-749b.json",
      "release_date": "2026-06-08",
      "date": "June 8, 2026",
      "name": "Macaron-V1-Preview-749B",
      "developer": "Mind Lab",
      "parameters": "749B total (744B base + 5B LoRA adapters) / 40B active",
      "context_window": "202,752 tokens",
      "license": "MIT",
      "key_features": "Official open-weight preview release in Mind Lab's Macaron V1 series, positioned as a personal-agent-focused post-trained model built on top of GLM-5.1 rather than a from-scratch frontier pretraining run. Release materials describe Macaron-V1-Preview-749B as a Mixture-of-LoRA architecture that combines a 744B GLM-5.1 base with five specialist 1B adapters for general chat, personal-agent workflows, coding and shell tasks, generative UI, and computer-agent behavior, with roughly 40B activated parameters at inference. The launch framing emphasizes explicit specialist routing for agent harness scenarios instead of a single opaque post-trained policy, aiming at multi-turn personal assistance where user goals and constraints evolve across the conversation. Mind Lab's related Macaron-A2UI research describes the broader 754B-class Macaron line as trained with parameter-efficient LoRA-based supervised fine-tuning followed by reward-driven reinforcement learning for personal-agent and generative-UI tasks, and public coverage of the Preview release highlights strong results on LivingBench, VitaBench, A2UI evaluation, PinchBench, and general math and coding tasks relative to contemporary open models. The release is also presented as unusually compute-efficient for its scale, with Mind Lab claiming development on fewer than 300 GPUs while keeping the checkpoint open under an MIT license.",
      "paper_url": "https://arxiv.org/abs/2605.24830",
      "links": [
        "https://huggingface.co/models?other=base_model%3Afinetune%3Azai-org%2FGLM-5.1",
        "https://macaron.im/mindlab",
        "https://huggingface.co/papers/2605.24830",
        "https://arxiv.org/abs/2605.24830",
        "https://finance.sina.com.cn/tech/roll/2026-06-08/doc-iniasmvs5567577.shtml",
        "https://aiweekly.co/node/2603"
      ],
      "page_url": "https://nlp.com.ai/models/macaron-v1-preview-749b/",
      "record_updated_at": "2026-10-11T08:37:45.631473+00:00"
    },
    {
      "slug": "nemotron-3-ultra-550b-a55b-base-bf16-bf16-nvfp4",
      "source_file": "2026-06-04-nemotron-3-ultra.json",
      "release_date": "2026-06-04",
      "date": "June 4, 2026",
      "name": "Nemotron 3 Ultra 550B-A55B (Base BF16, BF16, NVFP4)",
      "developer": "NVIDIA (USA)",
      "parameters": "550B total / 55B active (MoE)",
      "context_window": "1M tokens",
      "license": "OpenMDW-1.1",
      "key_features": "Official open-weight Nemotron 3 Ultra release family covering the base BF16 checkpoint, the post-trained BF16 checkpoint, and an inference-optimized NVFP4 quantized checkpoint for frontier-scale agentic reasoning. NVIDIA positions Nemotron 3 Ultra as the largest and most capable model in the Nemotron 3 line, built for long-running autonomous agents across coding, research, multi-step planning, tool use, synthesis, verification, and recovery workflows. The technical report describes a hybrid Mamba-Attention Mixture-of-Experts architecture with 108 total layers, 8,192 model dimension, 64 query heads, 2 KV heads, 512 experts per layer with top-22 activation, LatentMoE routing, and two shared-weight Multi-Token Prediction heads that enable native speculative decoding. NVIDIA says the model was pretrained in NVFP4 on 20T text tokens, then extended to 1M context, and post-trained with a pipeline combining supervised fine-tuning, unified RLVR across reasoning and agentic environments, and Multi-teacher On-Policy Distillation, with reasoning budget control exposed at inference time. Official release materials emphasize both accuracy and efficiency, including claims of 5.9x, 4.8x, and 1.6x higher inference throughput than GLM-5.1-754B-A40B, Kimi-K2.6-1T-A32B, and Qwen-3.5-397B-17B respectively on an 8K input / 64K output setting while maintaining on-par benchmark accuracy, plus open release of the checkpoints, training data, RL environments, and model recipes.",
      "paper_url": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf",
      "links": [
        "https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/",
        "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf",
        "https://developer.nvidia.com/topics/ai/nemotron",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4",
        "https://raw.githubusercontent.com/OpenMDW/OpenMDW/refs/heads/main/1.1/LICENSE.OpenMDW-1.1"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-3-ultra-550b-a55b-base-bf16-bf16-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.629658+00:00"
    },
    {
      "slug": "step-37-flash",
      "source_file": "2026-05-29-step-3.7-flash.json",
      "release_date": "2026-05-29",
      "date": "May 29, 2026",
      "name": "Step 3.7 Flash",
      "developer": "StepFun (China)",
      "parameters": "198B total / ~11B active (196B language + 1.8B vision)",
      "context_window": "256K tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight sparse MoE vision-language release positioned as a high-efficiency production model for real-world agents. StepFun describes Step 3.7 Flash as a multimodal model with a 196B language backbone plus a 1.8B vision encoder, activating about 11B parameters per token, with native image understanding, 256K context, and three selectable reasoning levels: low, medium, and high. The public model card emphasizes high-frequency agent workloads combining perception, search, reasoning, and tool orchestration, with claims of up to 400 tokens per second throughput and deployment support across vLLM, SGLang, Transformers, llama.cpp, NVIDIA NIM, and related infrastructure. Benchmark tables in the official materials highlight strong multimodal and agentic performance, including first place on SimpleVQA (Search), 67.1 on ClawEval-1.1, 49.5 on Toolathlon, 56.3 on SWE-Bench Pro, and 59.5 on Terminal-Bench 2.1. StepFun also presents the model as compatible with mainstream agent harnesses and suitable for multimodal enterprise workflows such as financial document parsing, visual search, and concurrent coding agents.",
      "paper_url": "",
      "links": [
        "https://static.stepfun.com/blog/step-3.7-flash/",
        "https://huggingface.co/stepfun-ai/Step-3.7-Flash",
        "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/main/README.md",
        "https://github.com/stepfun-ai/Step-3.7-Flash"
      ],
      "page_url": "https://nlp.com.ai/models/step-37-flash/",
      "record_updated_at": "2026-10-11T08:37:45.627289+00:00"
    },
    {
      "slug": "lfm25-8b-a1b-base-reasoning-tuned",
      "source_file": "2026-05-28-lfm2.5-8b-a1b.json",
      "release_date": "2026-05-28",
      "date": "May 28, 2026",
      "name": "LFM2.5-8B-A1B (Base, Reasoning-tuned)",
      "developer": "Liquid AI (USA)",
      "parameters": "8.3B total / 1.5B active (MoE)",
      "context_window": "128,000 tokens",
      "license": "LFM 1.0 license",
      "key_features": "Official open-weight LFM2.5 family release covering a pre-trained base checkpoint and a reasoning-tuned general-purpose checkpoint, positioned by Liquid AI as an on-device personal-assistant model line for tool use and complex instruction following on consumer hardware. Release materials say LFM2.5-8B-A1B builds on the earlier LFM2-8B-A1B with scaled pretraining from 12T to 38T tokens, large-scale reinforcement learning, an expanded context window from 32,768 to 128,000 tokens, and a doubled 128K vocabulary for better non-Latin tokenization efficiency. The published config identifies Liquid's hybrid sparse Lfm2MoeForCausalLM architecture with 24 layers, 32 experts, 4 experts active per token, 32 attention heads, 8 KV heads, 18 double-gated LIV convolution layers plus 6 GQA attention layers, and 128,000 maximum position embeddings. The Hugging Face card reports strong gains over LFM2-8B-A1B across instruction following, math, function calling, and hallucination-sensitive evaluations, including IFEval, IFBench, Multi-IF, MATH500, AIME25, BFCLv3, BFCLv4, and AA-Omniscience metrics. Liquid also emphasizes day-one deployment support across Transformers, llama.cpp, MLX, vLLM, SGLang, and ONNX, with the tuned checkpoint specifically recommended for agentic workflows, structured outputs, multilingual assistants, and on-device personal-assistant applications.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://www.liquid.ai/blog/lfm2-5-8b-a1b",
        "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B",
        "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-Base",
        "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B/raw/main/README.md",
        "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B/raw/main/config.json"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-8b-a1b-base-reasoning-tuned/",
      "record_updated_at": "2026-10-11T08:37:45.625016+00:00"
    },
    {
      "slug": "mellum2-12b-a25b-base-instruct-thinking",
      "source_file": "2026-05-26-mellum2-12b-a2.5b.json",
      "release_date": "2026-05-26",
      "date": "May 26, 2026",
      "name": "Mellum2-12B-A2.5B (Base, Instruct, Thinking)",
      "developer": "JetBrains",
      "parameters": "12B total / 2.5B active (MoE)",
      "context_window": "128K tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Mellum 2 family release covering base, instruct, and thinking checkpoints for software-engineering-focused text generation. JetBrains describes Mellum 2 as a 12B-parameter Mixture-of-Experts model with 2.5B active parameters per token, positioned as a successor to the earlier 4B dense Mellum and specialized for code generation and editing, debugging, multi-step reasoning, tool use and function calling, agentic coding, and conversational programming assistance. The Mellum2 Technical Report says the architecture uses 64 experts with 8 active per token, Grouped-Query Attention with 4 KV heads, Sliding Window Attention on three of every four layers, and a single Multi-Token Prediction head that also serves as a built-in draft model for speculative decoding. Training materials describe roughly 10.6T pretraining tokens, a three-phase curriculum that increasingly emphasizes curated code and math data, extension of the base model to a 128K context window via layer-selective YaRN, and a two-stage post-training pipeline of supervised fine-tuning followed by RLVR. The released variants separate direct-answer and explicit-reasoning behavior into Instruct and Thinking checkpoints respectively, and JetBrains also published same-day SFT intermediate checkpoints on Hugging Face alongside the main family release.",
      "paper_url": "https://arxiv.org/abs/2605.31268",
      "links": [
        "https://blog.jetbrains.com/ai/2026/06/mellum2-goes-open-source-a-fast-model-for-ai-workflows/",
        "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Base",
        "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Instruct",
        "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Thinking",
        "https://huggingface.co/api/models?author=JetBrains&search=Mellum2",
        "https://arxiv.org/abs/2605.31268"
      ],
      "page_url": "https://nlp.com.ai/models/mellum2-12b-a25b-base-instruct-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.622943+00:00"
    },
    {
      "slug": "laguna-m1",
      "source_file": "2026-05-26-laguna-m.1.json",
      "release_date": "2026-05-26",
      "date": "May 26, 2026",
      "name": "Laguna M.1",
      "developer": "Poolside (USA)",
      "parameters": "225.8B total / 23.4B active (MoE)",
      "context_window": "256K tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight flagship coding model in Poolside's Laguna family, positioned for long-horizon agentic software engineering and terminal work. Poolside describes Laguna M.1 as a sparse Mixture-of-Experts transformer trained completely from scratch in-house on more than 30T tokens using its internal 'Model Factory' stack, with 225.8B total parameters and 23.4B activated per token. The technical report and release blog emphasize that Laguna M.1 finished pretraining first and served as the foundation from which Poolside derived later Laguna-family lessons, while the open-weight release now includes post-trained and base checkpoints plus efficiency-oriented variants in the public Hugging Face collection. Official materials highlight large-scale pretraining on 6,144 interconnected NVIDIA Hopper GPUs, heavy synthetic-data use across the Laguna family, a distributed Muon optimizer implementation, and a fully asynchronous online agent-RL system built around Poolside's own coding-agent harness. Poolside reports benchmark results of 74.6 on SWE-bench Verified, 63.1 on SWE-bench Multilingual, 49.2 on SWE-Bench Pro, and 45.8 on Terminal-Bench 2.0 in the technical report, with all agentic evaluations run at 256K context using its internal harness and patched benchmark images to address infrastructure and reward-hacking issues. The Hugging Face collection and model cards also indicate broad open deployment support through base and post-trained checkpoints and multiple public formats, including BF16, FP8, and NVFP4 variants, under Apache 2.0 licensing.",
      "paper_url": "https://arxiv.org/abs/2605.27605",
      "links": [
        "https://poolside.ai/blog/laguna-a-deeper-dive",
        "https://huggingface.co/poolside/Laguna-M.1",
        "https://huggingface.co/poolside/Laguna-M.1-base",
        "https://huggingface.co/collections/poolside/laguna-m1",
        "https://arxiv.org/abs/2605.27605"
      ],
      "page_url": "https://nlp.com.ai/models/laguna-m1/",
      "record_updated_at": "2026-10-11T08:37:45.620783+00:00"
    },
    {
      "slug": "supra-50m-base-instruct",
      "source_file": "2026-05-21-supra-50m.json",
      "release_date": "2026-05-21",
      "date": "May 21, 2026",
      "name": "Supra-50M (Base, Instruct)",
      "developer": "SupraLabs",
      "parameters": "50M",
      "context_window": "1,024 tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Supra-50M family release covering both a base causal language model and an instruction-tuned variant, positioned by SupraLabs as the first model in its Scaling Up Plan. The base model card describes Supra-50M as a decoder-only LlamaForCausalLM architecture trained from scratch on 20B tokens from HuggingFaceFW/fineweb-edu, with a 32K custom byte-level BPE tokenizer, 12 layers, 512 hidden size, 1,408 intermediate size, 8 attention heads, 4 key-value heads with GQA, and 1,024 maximum position embeddings. Release materials claim the 50M model is competitive with or better than substantially larger small models on several benchmarks, including reported gains over GPT-2 124M on BLiMP, SciQ, ARC-Easy, and HellaSwag, while remaining CPU-friendly and tiny-model oriented. The instruct variant is built from the same pretrained base and supervised fine-tuned for four epochs on yahma/alpaca-cleaned, giving the family a chat-capable release alongside the raw base checkpoint. Both cards describe Apache 2.0 licensing and standard Transformers compatibility, with the family framed as a compact educational-web-text-trained English text-generation line under Project Chimera.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/SupraLabs/Supra-50M-Base",
        "https://huggingface.co/SupraLabs/Supra-50M-Instruct",
        "https://huggingface.co/SupraLabs/Supra-50M-Base/raw/main/README.md",
        "https://huggingface.co/SupraLabs/Supra-50M-Instruct/raw/main/README.md"
      ],
      "page_url": "https://nlp.com.ai/models/supra-50m-base-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.618296+00:00"
    },
    {
      "slug": "minicpm5-1b-final-sft",
      "source_file": "2026-05-21-minicpm5-1b.json",
      "release_date": "2026-05-21",
      "date": "May 21, 2026",
      "name": "MiniCPM5-1B (Final, SFT)",
      "developer": "OpenBMB",
      "parameters": "1,080,632,832",
      "context_window": "131,072 tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight MiniCPM5 family release covering the final MiniCPM5-1B checkpoint and the same-day SFT-only checkpoint that precedes RL and OPD post-training. OpenBMB positions MiniCPM5-1B as a dense 1B-class LlamaForCausalLM model for on-device and resource-constrained deployment, with native 131,072-token context, 24 layers, 16 query heads with 2 KV heads under GQA, built-in think and no-think chat modes through the same checkpoint, and native XML-style tool-calling support with SGLang parser integration. The release card claims 1B-class open-source SOTA within its comparison set, with especially visible gains in agentic tool use, code generation, and difficult reasoning versus small open models such as Qwen3 and Qwen3.5 think variants and LFM2.5-1.2B-Thinking. Training materials describe a three-stage stack of base training, mid-training, and post-training under OpenBMB's UltraData tiered data management framework, using Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math, and UltraData-SFT-2605. For the final checkpoint, the card highlights a post-training pipeline of SFT, specialized RL teachers, and On-Policy Distillation, with claimed average gains of 16 points on math, code, and instruction-following tasks while reducing max-token-budget overlong responses by 29 percentage points. The grouped SFT checkpoint represents the pre-RL and pre-OPD stage, while the final BF16 release is the fully post-trained model intended for local assistants, coding agents, tool-use workflows, and compact reasoning assistants.",
      "paper_url": "https://arxiv.org/pdf/2506.07900",
      "links": [
        "https://huggingface.co/openbmb/MiniCPM5-1B",
        "https://huggingface.co/openbmb/MiniCPM5-1B-SFT",
        "https://huggingface.co/openbmb/MiniCPM5-1B/raw/main/README.md",
        "https://huggingface.co/openbmb/MiniCPM5-1B-SFT/raw/main/README.md",
        "https://github.com/OpenBMB/MiniCPM"
      ],
      "page_url": "https://nlp.com.ai/models/minicpm5-1b-final-sft/",
      "record_updated_at": "2026-10-11T08:37:45.615925+00:00"
    },
    {
      "slug": "command-a-1",
      "source_file": "2026-05-20-command-a-plus.json",
      "release_date": "2026-05-20",
      "date": "May 20, 2026",
      "name": "Command A+",
      "developer": "Cohere (Canada)",
      "parameters": "25B active / 218B total (MoE)",
      "context_window": "128K input / 64K output",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Command A family release that Cohere positions as its fastest and most powerful language model yet, combining text, vision input, reasoning, translation, and agentic tool-use capabilities in one model. The Hugging Face model card describes Command A+ as a decoder-only sparse Mixture-of-Experts Transformer with 25B active parameters and 218B total parameters, using 128 experts with 8 active per token plus a shared expert. Cohere says the architecture interleaves sliding-window RoPE attention and global attention in a 3:1 ratio inherited from Command A, uses a fully dropless sparse MoE layer with token-choice routing, additive-bias load balancing, and a normalized-sigmoid top-k router. Release materials emphasize 48-language coverage, stronger agentic and tool-use performance than earlier Command-family models, and deployment across multiple official quantizations: BF16, FP8, and W4A4. The W4A4 card specifically highlights quantization-aware distillation and selective NVFP4 W4A4 quantization of MoE experts while keeping the attention path at full precision, targeting single-B200 or dual-H100 deployment with lower latency and higher throughput.",
      "paper_url": "",
      "links": [
        "https://cohere.com/blog/command-a-plus",
        "https://huggingface.co/CohereLabs/command-a-plus-05-2026-w4a4",
        "https://huggingface.co/CohereLabs/command-a-plus-05-2026-w4a4/raw/main/README.md",
        "https://docs.cohere.com/changelog"
      ],
      "page_url": "https://nlp.com.ai/models/command-a-1/",
      "record_updated_at": "2026-10-11T08:37:45.613407+00:00"
    },
    {
      "slug": "nemotron-labs-diffusion-3b-8b-14b-base-instruct-vlm-8b",
      "source_file": "2026-05-19-nemotron-labs-diffusion.json",
      "release_date": "2026-05-19",
      "date": "May 19, 2026",
      "name": "Nemotron-Labs-Diffusion (3B, 8B, 14B; Base + Instruct + VLM-8B)",
      "developer": "NVIDIA (United States)",
      "parameters": "3B / 8B / 14B dense LM family; VLM-8B is listed as 9B on Hugging Face",
      "context_window": "262,144 tokens",
      "license": "NVIDIA Nemotron Open Model License (LM variants); NSCLv1 (VLM-8B)",
      "key_features": "Official open-weight Nemotron tri-mode family that unifies three decoding modes inside one model: standard autoregressive decoding, diffusion-based parallel decoding, and self-speculation where diffusion drafts and autoregressive decoding verifies against a shared KV cache. NVIDIA positions the family as a dense 3B/8B/14B release spanning base, instruct, and vision-language variants, designed to adapt to different deployment regimes rather than forcing a single decoding strategy. The technical report says the models are trained with a joint autoregressive-plus-diffusion objective and a two-stage pipeline that first strengthens left-to-right priors and then enables joint AR and diffusion training; the VLM extends the same backbone with a Pixtral-style vision encoder. Official release claims emphasize both quality and decode efficiency versus open autoregressive and diffusion baselines, including roughly 6x more tokens per forward than Qwen3-8B at comparable accuracy for the 8B model, about 4x higher throughput on SPEED-Bench with SGLang on GB200, and stronger self-speculation acceptance and speed than Qwen3-8B-Eagle3. NVIDIA also highlights practical deployment support through Transformers, vLLM, SGLang, and Docker, plus an optional LoRA-enhanced drafter for higher self-speculation acceptance lengths.",
      "paper_url": "https://d1qx31qr3h6wln.cloudfront.net/publications/Nemotron_Diffusion_Tech_Report_v1.pdf?VersionId=db8_EMO8B.vmU26.jr7Le9pN3MqcUDNL",
      "links": [
        "https://research.nvidia.com/publication/2026-05_nemotron-labs-diffusion-tri-mode-language-model-unifying-autoregressive",
        "https://huggingface.co/collections/nvidia/nemotron-labs-diffusion",
        "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B",
        "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B",
        "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-14B",
        "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-VLM-8B",
        "https://d1qx31qr3h6wln.cloudfront.net/publications/Nemotron_Diffusion_Tech_Report_v1.pdf?VersionId=db8_EMO8B.vmU26.jr7Le9pN3MqcUDNL"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-labs-diffusion-3b-8b-14b-base-instruct-vlm-8b/",
      "record_updated_at": "2026-10-11T08:37:45.610433+00:00"
    },
    {
      "slug": "hrm-text-1b",
      "source_file": "2026-05-18-hrm-text-1b.json",
      "release_date": "2026-05-18",
      "date": "May 18, 2026",
      "name": "HRM-Text-1B",
      "developer": "Sapient Intelligence",
      "parameters": "1B",
      "context_window": "4096",
      "license": "Apache-2.0",
      "key_features": "Official open-weight 1B-parameter base language model built on Sapient Intelligence's Hierarchical Reasoning Model architecture. The release frames HRM-Text as an efficiency-focused from-scratch pretraining effort using a dual-timescale recurrent design with separate high-level and low-level transformer stacks that are reused across nested recurrence steps, aiming to increase effective compute depth without scaling parameter count in the usual way. Official model details describe PrefixLM training, gated attention with a per-head sigmoid output gate, parameterless Pre-RMSNorm, RoPE positional encoding, SwiGLU activations, a 65,536-token vocabulary, and a training run over 40B unique tokens. The model card explicitly positions this checkpoint as pre-alignment rather than an assistant model: it is non-chat, non-instruction-tuned, and intended as a starting point for further SFT or RL. Release materials also emphasize unusually low reported pretraining cost for the reference runs, with a 0.6B model trained on 8 H100s in about 50 hours and a 1B model on 16 H100s in about 46 hours, alongside benchmark claims including 84.7 on GSM8K, 56.5 on MATH, 82.3 on DROP, and 60.7 on MMLU for the 1B reference run.",
      "paper_url": "https://arxiv.org/abs/2605.20613",
      "links": [
        "https://huggingface.co/sapientinc/HRM-Text-1B",
        "https://github.com/sapientinc/HRM-Text",
        "https://huggingface.co/docs/transformers/main/en/model_doc/hrm_text"
      ],
      "page_url": "https://nlp.com.ai/models/hrm-text-1b/",
      "record_updated_at": "2026-10-11T08:37:45.607966+00:00"
    },
    {
      "slug": "equinox-31b",
      "source_file": "2026-05-18-equinox-31b.json",
      "release_date": "2026-05-18",
      "date": "May 18, 2026",
      "name": "Equinox-31B",
      "developer": "LatitudeGames",
      "parameters": "31B",
      "context_window": "262,144 tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Gemma 4 fine-tune positioned as a balanced longform roleplay and text-adventure model that sits between LatitudeGames' darker Wayfarer 2 style and the quieter slice-of-life Hearthfire style. The model card says Equinox-31B was trained with two epochs of supervised fine-tuning on top of Google Gemma 4 31B Instruct, using a balanced dataset that combines dark consequence-driven adventure data from Wayfarer 2 with longform character-focused Hearthfire data. Release materials emphasize second-person present-tense storytelling as the primary format, supplemented by a smaller third-person dataset in the same narrative style. The published config identifies a Gemma4ForConditionalGeneration architecture with 60 text layers, 5,376 hidden size, 21,504 intermediate size, 32 attention heads, 16 key-value heads, alternating sliding and full attention layers, and a 262,144-token text context window. LatitudeGames notes that no reasoning datasets were included during training, recommends Gemma 4's native chat template with thinking mode suppressed by default, and positions the model for narrative roleplay, interactive fiction, and game-master style responses. A separate GGUF quantization repo was later published for easier local deployment.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Equinox-31B",
        "https://huggingface.co/LatitudeGames/Equinox-31B/raw/main/README.md",
        "https://huggingface.co/LatitudeGames/Equinox-31B/raw/main/config.json",
        "https://huggingface.co/LatitudeGames/Equinox-31B-GGUF"
      ],
      "page_url": "https://nlp.com.ai/models/equinox-31b/",
      "record_updated_at": "2026-10-11T08:37:45.605494+00:00"
    },
    {
      "slug": "ring-26-1t",
      "source_file": "2026-05-15-ring-2.6-1t.json",
      "release_date": "2026-05-15",
      "date": "May 15, 2026",
      "name": "Ring-2.6-1T",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "1T",
      "context_window": "128K -> 256K (YaRN)",
      "license": "MIT",
      "key_features": "Official open-weight trillion-parameter Ring reasoning flagship aimed at real-world complex tasks, especially agent workflows, engineering development, scientific research analysis, complex business systems, and enterprise automation. The official model card positions Ring-2.6-1T as an execution-first model with stronger multi-step task decomposition, tool use, contextual planning, error correction, and long-horizon workflow stability than prior releases. InclusionAI highlights two adjustable reasoning-effort modes, high and xhigh, so developers can trade off speed, token overhead, and reasoning depth depending on task difficulty. The release also emphasizes an asynchronous reinforcement learning training pipeline combined with the IcePop algorithm to improve training efficiency and stability for trillion-parameter long-horizon RL. Official benchmark claims include 87.60 on PinchBench, 63.82 on ClawEval, and 95.32 on Tau2-Bench Telecom for agent execution, plus 66.18 on ARC-AGI-V2, 95.83 on AIME 26, and 88.27 on GPQA Diamond in xhigh reasoning mode. The official download table lists 128K context with expansion to 256K via YaRN, and the model card includes direct deployment guidance for Transformers, vLLM, Docker, and SGLang.",
      "paper_url": "https://arxiv.org/abs/2606.15079",
      "links": [
        "https://huggingface.co/inclusionAI/Ring-2.6-1T",
        "https://huggingface.co/collections/inclusionAI/ring-26",
        "https://modelscope.cn/models/inclusionAI/Ring-2.6-1T"
      ],
      "page_url": "https://nlp.com.ai/models/ring-26-1t/",
      "record_updated_at": "2026-10-11T08:37:45.603125+00:00"
    },
    {
      "slug": "intern-s2-preview",
      "source_file": "2026-05-15-intern-s2-preview.json",
      "release_date": "2026-05-15",
      "date": "May 15, 2026",
      "name": "Intern-S2-Preview",
      "developer": "Shanghai AI Laboratory / InternLM (China)",
      "parameters": "35B-A3B-class MoE; 36B total parameters in Hugging Face metadata",
      "context_window": "256K tokens native; evaluations capped at 128K text / 64K multimodal",
      "license": "Apache-2.0",
      "key_features": "Official open-weight scientific multimodal foundation model from the InternLM team, positioned as an efficient successor for scientific reasoning, scientific generation, and workflow automation. Intern-S2-Preview-35B is a roughly 35B-total, 3B-active Mixture-of-Experts model continued-pretrained from Qwen3.5-35B. Its training emphasizes task scaling across hundreds of professional scientific tasks rather than parameter scaling alone, aiming to match the trillion-scale Intern-S1-Pro on multiple core scientific tasks with far fewer active parameters. The later Intern-S2 technical report documents a unified pipeline spanning scientific multimodal pre-training, supervised fine-tuning, scalable multi-task reinforcement learning, black- and white-box agentic RL, and on-policy distillation, supported by partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware agent-experience assembly. Release materials emphasize stronger multimodal scientific reasoning, improved scientific-agent capabilities, enhanced spatial modeling for small-molecule structures, and a real-valued prediction module; the team also describes it as the first open-source model combining material crystal-structure generation with strong general capabilities. Shared-weight Multi-Token Prediction with KL loss and chain-of-thought compression are used to improve generation speed and reasoning brevity while preserving capability. The model supports native tool calling and deployment through LMDeploy, vLLM, and SGLang. Official API documentation lists a 256K context window, while published evaluations cap text reasoning at 128K and multimodal evaluation at 64K.",
      "paper_url": "https://arxiv.org/abs/2608.13505",
      "links": [
        "https://huggingface.co/internlm/Intern-S2-Preview",
        "https://huggingface.co/internlm/Intern-S2-Preview-FP8",
        "https://huggingface.co/collections/internlm/intern-s2",
        "https://github.com/InternLM/Intern-S1",
        "https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview",
        "https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview-FP8"
      ],
      "page_url": "https://nlp.com.ai/models/intern-s2-preview/",
      "record_updated_at": "2026-10-11T08:37:45.600649+00:00"
    },
    {
      "slug": "zaya1-8b-zaya1-8b-zaya1-reasoning-base",
      "source_file": "2026-05-06-zaya1-8b.json",
      "release_date": "2026-05-06",
      "date": "May 6, 2026",
      "name": "ZAYA1-8B (ZAYA1-8B + ZAYA1-reasoning-base)",
      "developer": "Zyphra (USA)",
      "parameters": "8.4B total / 760M active (MoE)",
      "context_window": "Unknown",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Zyphra release for compact high-efficiency reasoning, math, and coding workloads, published alongside public cloud access on Zyphra Cloud. The Hugging Face model card describes ZAYA1-8B as a small mixture-of-experts language model trained end-to-end by Zyphra, with under 1B active parameters and strong long-form reasoning performance relative to much larger open-weight models. Zyphra's release post says it is the first MoE model pretrained, midtrained, and supervised fine-tuned on an AMD Instinct MI300 stack, using a 1,024-node MI300X cluster with AMD Pensando Pollara interconnect on infrastructure built with IBM. Release materials attribute its efficiency to Compressed Convolutional Attention, an MLP-based expert router, and learned residual scaling, then describe a five-stage post-training pipeline covering SFT, reasoning warmup, RLVE-Gym training, large-scale math-and-code RL, and a final RLHF/RLAIF pass. Zyphra also introduces Markovian RSA as a test-time-compute method co-trained with the model, claiming major gains on difficult math benchmarks under larger reasoning budgets and positioning ZAYA1-8B as competitive with substantially larger reasoning models on AIME'26, HMMT Feb.'26, LiveCodeBench v6, GPQA-Diamond, and MMLU-Pro.",
      "paper_url": "https://www.zyphra.com/zaya1-8b-technical-report",
      "links": [
        "https://www.zyphra.com/post/zaya1-8b",
        "https://www.zyphra.com/zaya1-8b-technical-report",
        "https://huggingface.co/Zyphra/ZAYA1-8B",
        "https://huggingface.co/Zyphra/ZAYA1-reasoning-base",
        "https://cloud.zyphra.com/"
      ],
      "page_url": "https://nlp.com.ai/models/zaya1-8b-zaya1-8b-zaya1-reasoning-base/",
      "record_updated_at": "2026-10-11T08:37:45.598086+00:00"
    },
    {
      "slug": "mistral-medium-35-128b-eagle",
      "source_file": "2026-04-29-mistral-medium-3.5.json",
      "release_date": "2026-04-29",
      "date": "April 29, 2026",
      "name": "Mistral Medium 3.5 (128B + EAGLE)",
      "developer": "Mistral AI (France)",
      "parameters": "128B (dense)",
      "context_window": "256K tokens",
      "license": "Modified MIT License",
      "key_features": "Official open-weight flagship merged Mistral release that unifies instruction-following, reasoning, coding, and vision in a single dense 128B model. Mistral positions Medium 3.5 as the successor to both Mistral Medium 3.1 and Magistral in Le Chat, and as the replacement for Devstral 2 in Vibe. Release materials emphasize configurable reasoning effort per request, multimodal text-plus-image input with text output, strong system-prompt adherence, native function calling and JSON output, and self-hosting on as few as four GPUs. The official model card reports 77.6% on SWE-Bench Verified and 91.4 on tau-cubed Telecom, with the model aimed at long-horizon coding, agentic tool use, and structured productivity workflows. Mistral also released a companion EAGLE head for faster speculative decoding in vLLM deployments.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5",
        "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B",
        "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B-EAGLE"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-medium-35-128b-eagle/",
      "record_updated_at": "2026-10-11T08:37:45.595499+00:00"
    },
    {
      "slug": "ling-26-1t",
      "source_file": "2026-04-29-ling-2.6-1t.json",
      "release_date": "2026-04-29",
      "date": "April 29, 2026",
      "name": "Ling-2.6-1T",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "1T",
      "context_window": "262,144 tokens",
      "license": "MIT",
      "key_features": "Official open-weight trillion-parameter Ling flagship for complex coding, multi-step execution, and agentic workflows. The Hugging Face model card describes Ling-2.6-1T as a comprehensive flagship model optimized for real-world complex scenarios, using a hybrid MLA plus Linear Attention architecture to reduce long-context latency and VRAM costs while improving throughput. InclusionAI highlights a post-training 'Fast Thinking' strategy based on Contextual Process Redundancy Suppression to reduce verbose chain-of-thought overhead and lower output-token costs while maintaining high capability. Release materials claim open-source state-of-the-art performance on execution-heavy benchmarks including AIME26, SWE-bench Verified, BFCL-V4, TAU2-Bench, and IFBench, with strong long-context performance on MRCR from 16K to 256K. The model is positioned as production-ready for engineering and enterprise agent workflows, with explicit compatibility claims for frameworks such as Claude Code, OpenClaw, OpenCode, and CodeBuddy, plus deployment guidance for both standard inference and MTP-accelerated serving in SGLang and vLLM.",
      "paper_url": "https://arxiv.org/abs/2606.15079",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-2.6-1T",
        "https://openrouter.ai/inclusionai/ling-2.6-1t/providers"
      ],
      "page_url": "https://nlp.com.ai/models/ling-26-1t/",
      "record_updated_at": "2026-10-11T08:37:45.593121+00:00"
    },
    {
      "slug": "granite-41-3b-8b-30b-base-instruct",
      "source_file": "2026-04-29-granite-4.1.json",
      "release_date": "2026-04-29",
      "date": "April 29, 2026",
      "name": "Granite 4.1 (3B, 8B, 30B; Base + Instruct)",
      "developer": "IBM (USA)",
      "parameters": "3B, 8B, 30B",
      "context_window": "128K tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight Granite 4.1 family release spanning 3B, 8B, and 30B text models, each published in both base and chat-capable variants. Public Hugging Face configs for the family indicate dense decoder-only GraniteForCausalLM architecture with 131,072 maximum position embeddings, Granite chat templating for conversational and tool-oriented prompting, and Apache 2.0 licensing across the release. IBM positions Granite as an enterprise-oriented model family emphasizing practical instruction following, deployment flexibility, and compatibility with standard open-source inference stacks. The 4.1 release extends that line with multiple scale points so teams can trade off quality, latency, and serving cost while staying within one architecture family and prompt format.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/blog/ibm-granite/granite-4-1",
        "https://huggingface.co/ibm-granite/granite-4.1-3b",
        "https://huggingface.co/ibm-granite/granite-4.1-8b",
        "https://huggingface.co/ibm-granite/granite-4.1-30b",
        "https://huggingface.co/ibm-granite/granite-4.1-3b-base",
        "https://huggingface.co/ibm-granite/granite-4.1-8b-base",
        "https://huggingface.co/ibm-granite/granite-4.1-30b-base"
      ],
      "page_url": "https://nlp.com.ai/models/granite-41-3b-8b-30b-base-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.590601+00:00"
    },
    {
      "slug": "talkie-13b-talkie-1930-13b-base-talkie-1930-13b-it-talkie-web-13b-base",
      "source_file": "2026-04-28-talkie-13b.json",
      "release_date": "2026-04-28",
      "date": "April 28, 2026",
      "name": "Talkie-13B (talkie-1930-13b-base, talkie-1930-13b-it, talkie-web-13b-base)",
      "developer": "talkie",
      "parameters": "13B",
      "context_window": "Unknown",
      "license": "Apache 2.0",
      "key_features": "Open-weight 13B language model family designed for historical-language and temporal-generalization research. The flagship vintage checkpoints are trained on 260B tokens of pre-1931 English-language text, creating a strict historical knowledge cutoff intended to reduce modern-data contamination and enable controlled experiments on what models can infer without exposure to later events, web text, or code-heavy modern corpora. The family includes a vintage base model, an instruction-tuned vintage post-train, and a matched modern web base model trained on FineWeb with identical architecture and training FLOPs for controlled comparisons between historical and contemporary pretraining data. The instruction-tuned model was fine-tuned on instruction-response pairs extracted from pre-1931 reference works such as etiquette manuals, encyclopedias, and letter-writing manuals, then further improved with online DPO using an LLM-as-a-judge. Release materials position Talkie as a testbed for historical reasoning, style transfer, temporal surprise analysis, and contamination-free evaluation rather than as a frontier general-purpose assistant.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/collections/talkie-lm/talkie-13b",
        "https://huggingface.co/talkie-lm/talkie-1930-13b-base",
        "https://huggingface.co/talkie-lm/talkie-1930-13b-it",
        "https://huggingface.co/talkie-lm/talkie-web-13b-base",
        "https://github.com/talkie-lm/talkie"
      ],
      "page_url": "https://nlp.com.ai/models/talkie-13b-talkie-1930-13b-base-talkie-1930-13b-it-talkie-web-13b-base/",
      "record_updated_at": "2026-10-11T08:37:45.587978+00:00"
    },
    {
      "slug": "sensenova-u1-8b-mot-8b-mot-sft",
      "source_file": "2026-04-28-sensenova-u1.json",
      "release_date": "2026-04-28",
      "date": "April 28, 2026",
      "name": "SenseNova U1 (8B-MoT + 8B-MoT-SFT)",
      "developer": "SenseNova / SenseTime (China)",
      "parameters": "8B MoT",
      "context_window": "32K tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight native multimodal SenseNova release built to unify multimodal understanding, reasoning, and generation inside a single monolithic architecture rather than using separate visual encoders or VAEs. The model card presents SenseNova U1 as a NEO-Unify-based system that models language and visual information end to end as a unified compound, supports image understanding, image generation, image editing, and interleaved image-text generation, and is intended as a step toward natively multimodal agentic learning. The currently public open release consists of two 8B Lite checkpoints: a base 8B-MoT model and an 8B-MoT-SFT model. Release materials claim open-source state-of-the-art performance across multimodal understanding and generation benchmarks, strong dense information rendering for posters, presentations, comics, resumes, and other infographic-style outputs, and native interleaved generation for illustrated tutorials and visually grounded long-form content. The card also notes a current visual understanding context limit of 32K tokens and describes the SFT models as trained through understanding warmup, generation pre-training, unified mid-training, unified SFT, and an initial round of text-to-image RL training.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/sensenova/SenseNova-U1-8B-MoT",
        "https://huggingface.co/sensenova/SenseNova-U1-8B-MoT-SFT"
      ],
      "page_url": "https://nlp.com.ai/models/sensenova-u1-8b-mot-8b-mot-sft/",
      "record_updated_at": "2026-10-11T08:37:45.585417+00:00"
    },
    {
      "slug": "nemotron-3-nano-omni-reasoning-bf16-fp8-nvfp4",
      "source_file": "2026-04-28-nemotron-3-nano-omni.json",
      "release_date": "2026-04-28",
      "date": "April 28, 2026",
      "name": "Nemotron 3 Nano Omni (Reasoning; BF16 + FP8 + NVFP4)",
      "developer": "NVIDIA (USA)",
      "parameters": "31B total / 3B active (30B-A3B class MoE)",
      "context_window": "256K tokens",
      "license": "NVIDIA Open Model Agreement",
      "key_features": "Official open-weight multimodal Nemotron release for enterprise-grade omni-understanding across video, audio, image, and text inputs with text output. NVIDIA positions Nemotron 3 Nano Omni as a native multimodal reasoning model for document intelligence, meeting and media analysis, OCR-heavy workflows, transcription, and GUI automation for agentic applications. The model card describes a Mamba2-Transformer hybrid Mixture-of-Experts architecture built on the Nemotron 3 Nano 30B-A3B backbone, extended with a CRADIO v4-H vision encoder and Parakeet speech encoder. Release materials emphasize integrated video plus speech comprehension, OCR reasoning, GUI understanding, JSON output, tool calling, word-level timestamps for transcription, and configurable reasoning output. The Hugging Face card lists support for up to 2-minute video inputs, audio files up to 1 hour, and maximum context length up to 256K tokens, with BF16, FP8, and NVFP4 public weight releases published together. NVIDIA also highlights enterprise deployment support through vLLM, TensorRT-LLM, SGLang, llama.cpp, and Ollama-oriented workflows.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16",
        "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8",
        "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4",
        "https://developer.nvidia.com/blog/building-nvidia-nemotron-3-agents-for-reasoning-multimodal-rag-voice-and-safety/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-3-nano-omni-reasoning-bf16-fp8-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.582862+00:00"
    },
    {
      "slug": "ling-26-flash",
      "source_file": "2026-04-28-ling-2.6-flash.json",
      "release_date": "2026-04-28",
      "date": "April 28, 2026",
      "name": "Ling-2.6-flash",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "104B total / 7.4B active (MoE)",
      "context_window": "256K tokens (262,144)",
      "license": "Unknown",
      "key_features": "Open-weight instant Ling release focused on high-throughput agentic workloads, long-context document processing, and code-heavy execution. Official materials describe Ling-2.6-flash as a sparse Mixture-of-Experts model with 104B total parameters and 7.4B activated parameters per token, positioned as a cost-efficient successor in the Ling family with full tool-calling and agent collaboration support. Ant's documentation highlights 256K context length, strong long-range retrieval across full-context positions, and production-oriented inference efficiency advantages from low activated parameter count. Third-party launch materials and ecosystem listings describe the model as optimized for token efficiency and agent reinforcement learning, with benchmark claims including strong results on agent-oriented evaluations such as BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval, and PinchBench, while maintaining competitive general intelligence and coding performance for its active compute budget.",
      "paper_url": "https://arxiv.org/abs/2606.15079",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-2.6-flash",
        "https://developer.ant-ling.com/en/docs/models/ling",
        "https://developer.ant-ling.com/en/docs/tutorials/sampling"
      ],
      "page_url": "https://nlp.com.ai/models/ling-26-flash/",
      "record_updated_at": "2026-10-11T08:37:45.580297+00:00"
    },
    {
      "slug": "laguna-xs2",
      "source_file": "2026-04-28-laguna-xs.2.json",
      "release_date": "2026-04-28",
      "date": "April 28, 2026",
      "name": "Laguna XS.2",
      "developer": "Poolside (USA)",
      "parameters": "33B total / 3B active (MoE)",
      "context_window": "128K tokens",
      "license": "Apache 2.0",
      "key_features": "Poolside's first open-weight model and second-generation XS-class agentic coding release, trained from scratch as a Mixture-of-Experts model for long-horizon software work. Official release materials describe Laguna XS.2 as a 33B-parameter model with 3B activated parameters trained on more than 30T tokens, incorporating lessons from Laguna M.1 across data, synthetic data generation, and agent reinforcement learning. Poolside reports 44.5% on SWE-bench Pro, 68.2% on SWE-bench Verified, 62.4% on SWE-bench Multilingual, and 30.1% on Terminal-Bench 2.0, positioning it as a strong open-weight model in its size class for coding agents. The blog also highlights a training stack built around Poolside's internal Titan codebase, distributed Muon optimizer, asynchronous long-horizon agent RL, and substantial synthetic-data use, with synthetic data contributing about 13% of the final XS.2 training mix and the Laguna family using roughly 4.4T+ synthetic tokens overall. Release materials further note day-one TensorRT-LLM support, an NVFP4 version for NVIDIA Blackwell, and compatibility with Poolside's released agent harness runtime.",
      "paper_url": "https://huggingface.co/poolside/Laguna-XS.2/resolve/main/laguna.pdf",
      "links": [
        "https://poolside.ai/blog/laguna-a-deeper-dive",
        "https://openrouter.ai/provider/poolside",
        "https://huggingface.co/poolside/Laguna-XS.2",
        "https://huggingface.co/poolside/Laguna-XS.2/resolve/main/laguna.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/laguna-xs2/",
      "record_updated_at": "2026-10-11T08:37:45.577720+00:00"
    },
    {
      "slug": "mimo-v25-mimo-v25-mimo-v25-base-mimo-v25-pro-mimo-v25-pro-base",
      "source_file": "2026-04-27-mimo-v2.5-series.json",
      "release_date": "2026-04-27",
      "date": "April 27, 2026",
      "name": "MiMo-V2.5 (MiMo-V2.5, MiMo-V2.5-Base, MiMo-V2.5-Pro, MiMo-V2.5-Pro-Base)",
      "developer": "Xiaomi MiMo (China)",
      "parameters": "MiMo-V2.5: 310B total / 15B active (Sparse MoE); MiMo-V2.5-Pro: 1.02T total / 42B active (MoE)",
      "context_window": "1M tokens (Base variants: 256K)",
      "license": "MIT",
      "key_features": "Official open-source MiMo-V2.5 family release covering both the native omnimodal MiMo-V2.5 track and the larger MiMo-V2.5-Pro agentic track, each published with Base and full post-trained variants. Xiaomi positions MiMo-V2.5 as a unified text-image-video-audio model built on the MiMo-V2-Flash backbone with dedicated in-house vision and audio encoders, 310B total parameters with 15B active, roughly 48T training tokens, and a 1M-token context window. The Hugging Face model card reports a 729M vision encoder, a 261M audio encoder, and three Multi-Token Prediction layers. Xiaomi positions MiMo-V2.5-Pro as the flagship long-horizon coding and agent model, with 1.02T total parameters, 42B active parameters, hybrid attention, three Multi-Token Prediction layers, 27T pretraining tokens, and a 1M-token context window. Both tracks use staged training with supervised fine-tuning, agentic reinforcement learning, and Multi-Teacher On-Policy Distillation. Release materials emphasize frontier-level open-source performance in agentic workflows, long-context coherence, coding, and multimodal understanding; Xiaomi cites Claw-Eval gains, strong token efficiency, and autonomous runs spanning hundreds to thousands of tool calls, including a SysY compiler task completed with a perfect hidden-test score and a full video editor generated end to end.",
      "paper_url": "",
      "links": [
        "https://mimo.xiaomi.com/mimo-v2-5",
        "https://mimo.xiaomi.com/mimo-v2-5-pro",
        "https://x.com/XiaomiMiMo/status/2048821516079661561",
        "https://huggingface.co/collections/XiaomiMiMo/mimo-v25",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.5",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Base",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-Base"
      ],
      "page_url": "https://nlp.com.ai/models/mimo-v25-mimo-v25-mimo-v25-base-mimo-v25-pro-mimo-v25-pro-base/",
      "record_updated_at": "2026-10-11T08:37:45.575046+00:00"
    },
    {
      "slug": "hy3-preview-instruct-base",
      "source_file": "2026-04-23-hy3-preview.json",
      "release_date": "2026-04-23",
      "date": "April 23, 2026",
      "name": "Hy3-preview (Instruct + Base)",
      "developer": "Tencent Hy Team (China)",
      "parameters": "295B total / 21B active (MoE) + 3.8B MTP layer",
      "context_window": "256K tokens (262,144)",
      "license": "Tencent Hy Community License Agreement",
      "key_features": "Officially open-sourced on April 23, 2026, Hy3-preview is positioned by Tencent as their strongest shipped model so far and the first trained on a rebuilt infrastructure. Release materials emphasize major gains in complex reasoning, instruction following, context learning, coding, and agent tasks, with strongest claimed improvements in Code & Agent benchmarks (including SWE-bench Verified, Terminal-Bench 2.0, BrowseComp, and WideSearch). The model card describes a large MoE architecture and notes deployable chat-completions usage with reasoning-effort controls (no_think/low/high), while also releasing a paired pre-trained Base checkpoint under the same release.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/tencent/Hy3-preview",
        "https://huggingface.co/tencent/Hy3-preview-Base",
        "https://huggingface.co/tencent/Hy3-preview/raw/main/README.md",
        "https://huggingface.co/tencent/Hy3-preview/raw/main/config.json",
        "https://huggingface.co/api/models/tencent/Hy3-preview",
        "https://modelscope.cn/models/Tencent-Hunyuan/Hy3-preview"
      ],
      "page_url": "https://nlp.com.ai/models/hy3-preview-instruct-base/",
      "record_updated_at": "2026-10-11T08:37:45.572541+00:00"
    },
    {
      "slug": "deepseek-v4-flash-pro-base-instruct",
      "source_file": "2026-04-23-deepseek-v4.json",
      "release_date": "2026-04-23",
      "date": "April 23, 2026",
      "name": "DeepSeek-V4 (Flash + Pro; Base + Instruct)",
      "developer": "DeepSeek AI (China)",
      "parameters": "Flash: 284B total / 13B active (MoE); Pro: 1.6T total / 49B active (MoE)",
      "context_window": "1,000,000 tokens",
      "license": "MIT",
      "key_features": "Official DeepSeek-V4 preview release with open weights for both Flash and Pro tracks, each provided in Base and Instruct variants. Release materials emphasize million-token context efficiency via hybrid attention upgrades, stronger long-context reasoning, and expanded coding/agent capability. DeepSeek positions V4-Pro-Max as its frontier open-source reasoning mode and claims top-tier open-source performance with narrowed gaps to leading closed models on difficult reasoning and agent benchmarks. Model pages provide local deployment guidance and multi-mode reasoning controls (Non-think, Think, Think Max).",
      "paper_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf",
      "links": [
        "https://huggingface.co/collections/deepseek-ai/deepseek-v4",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Base",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-Base",
        "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf",
        "https://api-docs.deepseek.com/"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v4-flash-pro-base-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.569983+00:00"
    },
    {
      "slug": "qwen36-27b-base-fp8",
      "source_file": "2026-04-22-qwen3.6-27b.json",
      "release_date": "2026-04-22",
      "date": "April 22, 2026",
      "name": "Qwen3.6-27B (Base, FP8)",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "27B dense; FP8 variant released",
      "context_window": "262,144 tokens (native; extensible to ~1,010,000)",
      "license": "Apache-2.0",
      "key_features": "Officially presented as an open-weight dense 27B multimodal model delivering flagship-level agentic coding at a practical deployment scale. Release messaging claims Qwen3.6-27B surpasses the previous open-source flagship Qwen3.5-397B-A17B across major coding benchmarks while retaining strong text and multimodal reasoning. The model card highlights default thinking mode plus non-thinking mode support, and includes preserve_thinking to retain historical reasoning traces for iterative agent workflows. Qwen reports native 262K context support with long-context extension guidance up to about 1.01M tokens, with deployment paths across Transformers, vLLM, SGLang, and related stacks.",
      "paper_url": "",
      "links": [
        "https://qwen.ai/blog?id=qwen3.6-27b",
        "https://huggingface.co/Qwen/Qwen3.6-27B",
        "https://huggingface.co/Qwen/Qwen3.6-27B-FP8",
        "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/README.md",
        "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json",
        "https://huggingface.co/api/models/Qwen/Qwen3.6-27B"
      ],
      "page_url": "https://nlp.com.ai/models/qwen36-27b-base-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.567454+00:00"
    },
    {
      "slug": "kimi-k26",
      "source_file": "2026-04-20-kimi-k2.6.json",
      "release_date": "2026-04-20",
      "date": "April 20, 2026",
      "name": "Kimi-K2.6",
      "developer": "Moonshot AI (China)",
      "parameters": "1T total / 32B active (MoE)",
      "context_window": "256K tokens (262,144)",
      "license": "Modified MIT License",
      "key_features": "Officially announced as an open-source Kimi release focused on long-horizon coding, coding-driven design, agent swarm orchestration, and proactive autonomous execution. Release materials position K2.6 as open-source SOTA on multiple agentic and coding benchmarks (including HLE with tools, SWE-Bench Pro, SWE-Bench Multilingual, BrowseComp, and Toolathlon) and highlight sustained multi-step tool use for extended engineering workflows. Model card summary reports K2.6 as a native multimodal agentic model and documents 1T total / 32B activated MoE architecture with 256K context length and Modified MIT licensing. Public model availability appears to have started as a Hugging Face upload on April 14, 2026, with official public release communications published on April 20, 2026.",
      "paper_url": "",
      "links": [
        "https://www.kimi.com/blog/kimi-k2-6",
        "https://x.com/Kimi_Moonshot/status/2046249571882500354",
        "https://huggingface.co/moonshotai/Kimi-K2.6",
        "https://huggingface.co/moonshotai/Kimi-K2.6/raw/main/README.md",
        "https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/LICENSE",
        "https://huggingface.co/api/models/moonshotai/Kimi-K2.6"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k26/",
      "record_updated_at": "2026-10-11T08:37:45.565071+00:00"
    },
    {
      "slug": "qwen36-35b-a3b-base-fp8",
      "source_file": "2026-04-15-qwen3.6-35b-a3b.json",
      "release_date": "2026-04-15",
      "date": "April 15, 2026",
      "name": "Qwen3.6-35B-A3B (Base, FP8)",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "35B total / 3B active (MoE); FP8 variant released",
      "context_window": "262,144 tokens (native; extensible to ~1,010,000)",
      "license": "Apache-2.0",
      "key_features": "Open-weight Qwen3.6 release positioned as the first open variant in the Qwen3.6 line, focused on stability and real-world coding productivity. Official release notes emphasize stronger agentic coding behavior for frontend and repository-level workflows, plus a new thinking-preservation option that retains historical reasoning context to improve iterative development. The architecture remains a sparse MoE design with 256 experts and 8 routed experts plus shared capacity, combining Gated DeltaNet and gated attention blocks in a hybrid layout tuned for coding and agent tasks. Published benchmark tables in release materials report gains over prior Qwen3.5-35B-A3B on multiple coding-agent evaluations including SWE-Bench Pro, Terminal-Bench 2.0, Claw-Eval averages, NL2Repo, and QwenWebBench, while maintaining competitive general-agent and knowledge performance. Release artifacts are provided in standard Transformers format with compatibility guidance for vLLM, SGLang, KTransformers, and related inference stacks.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8",
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/README.md",
        "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json",
        "https://qwen.ai/blog?id=qwen3.6-35b-a3b"
      ],
      "page_url": "https://nlp.com.ai/models/qwen36-35b-a3b-base-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.562988+00:00"
    },
    {
      "slug": "minimax-m27",
      "source_file": "2026-04-09-minimax-m2.7.json",
      "release_date": "2026-04-09",
      "date": "April 9, 2026",
      "name": "MiniMax-M2.7",
      "developer": "MiniMax (China)",
      "parameters": "~228.7B total (MoE; 256 experts, 8 active per token; FP8 quantization)",
      "context_window": "196,608 tokens",
      "license": "Non-commercial MIT-style license (commercial use requires prior written authorization from MiniMax; includes attribution and use restrictions)",
      "key_features": "Open-weight MiniMax-M2.7 release positioned as a flagship agentic engineering model with strong software and productivity workflow performance. Official release materials frame M2.7 as a model that deeply participates in its own evolution, describing self-improvement loops over 100+ rounds with reported internal gains, and emphasize long-horizon task execution with Agent Teams, skill composition, and dynamic tool usage. Benchmark claims in release docs highlight strong scores across SWE-Pro, SWE Multilingual, Multi SWE Bench, VIBE-Pro, Terminal Bench 2, NL2Repo, Toolathon, GDPval-AA, and MM Claw evaluations, alongside practical production positioning for incident analysis and recovery workflows. Deployment guidance is provided for SGLang, vLLM, and Transformers, with official API and agent platform support. The later MiniMax-M2 series paper presents M2.7 as the latest checkpoint in a Mixture-of-Experts family built for agentic deployment, with the flagship M2 architecture at 229.9B total parameters and 9.8B activated per token, large-scale verifiable agent trajectories, artifact-aligned rewards, and Forge as the underlying agent-native RL system; it also specifically highlights M2.7 as an early step toward self-evolution through autonomous debugging and scaffold modification.",
      "paper_url": "https://arxiv.org/abs/2605.26494",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M2.7",
        "https://www.minimax.io/news/minimax-m27-en",
        "https://github.com/MiniMax-AI/MiniMax-M2.7",
        "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json",
        "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE",
        "https://arxiv.org/abs/2605.26494"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m27/",
      "record_updated_at": "2026-10-11T08:37:45.560665+00:00"
    },
    {
      "slug": "lfm25-vl-450m",
      "source_file": "2026-04-08-lfm2.5-vl-450m.json",
      "release_date": "2026-04-08",
      "date": "April 8, 2026",
      "name": "LFM2.5-VL-450M",
      "developer": "Liquid AI (USA)",
      "parameters": "450M-class VLM (LFM2.5-350M text backbone + SigLIP2 NaFlex ~86M vision encoder)",
      "context_window": "128K tokens",
      "license": "LFM Open License v1.0",
      "key_features": "Open-weight refreshed 450M vision-language release built on the LFM2.5-350M backbone and positioned for stronger real-world edge multimodal performance than LFM2-VL-450M. Official model-card claims highlight improved instruction following on vision-language tasks, improved multilingual vision understanding across nine languages, and new/expanded grounded capabilities including bounding-box prediction and object detection, plus function-calling support for text-only prompts. Release documentation emphasizes practical on-device deployment and flexibility: native-resolution handling up to 512x512 without forced aspect-ratio distortion, tiling with thumbnail context for larger images, and tunable image-token budgets at inference for speed/quality trade-offs. Published benchmark tables report broad gains over LFM2-VL-450M across MMStar, RealWorldQA, MMBench, MM-IFEval, OCRBench, CountBench, RefCOCO-M, and language-side IFEval/Multi-IF/BFCLv4, with release availability across Transformers, GGUF/llama.cpp, ONNX, and MLX variants.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2.5-VL-450M",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-450M/raw/main/README.md",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-450M/raw/main/config.json",
        "https://www.liquid.ai/blog/lfm2-5-vl-450m",
        "https://docs.liquid.ai/lfm/getting-started/welcome",
        "https://arxiv.org/abs/2511.23404"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-vl-450m/",
      "record_updated_at": "2026-10-11T08:37:45.558534+00:00"
    },
    {
      "slug": "glm-51-base-fp8",
      "source_file": "2026-04-07-glm-5.1.json",
      "release_date": "2026-04-07",
      "date": "April 7, 2026",
      "name": "GLM-5.1 (Base, FP8)",
      "developer": "Z.AI / Zhipu AI (China)",
      "parameters": "~753.9B total (MoE-DSA family; FP8 variant also released)",
      "context_window": "202,752 tokens",
      "license": "MIT",
      "key_features": "Open-weight GLM-5.1 release positioned as Z.AI's next-generation flagship for agentic engineering, with official materials claiming significantly stronger coding capability than GLM-5. The model card reports state-of-the-art performance on SWE-Bench Pro and sizable gains over GLM-5 on NL2Repo and Terminal-Bench 2.0, while also improving across broader tool-use and web-browsing evaluations in the published benchmark table. Release messaging emphasizes long-horizon agent reliability rather than only first-pass quality, claiming GLM-5.1 can sustain iterative optimization over hundreds of rounds and thousands of tool calls by revising strategy based on intermediate results. The release includes both a standard BF16 checkpoint and an FP8 variant, with local serving guidance across SGLang, vLLM, xLLM, Transformers, and KTransformers.",
      "paper_url": "https://arxiv.org/abs/2602.15763",
      "links": [
        "https://huggingface.co/zai-org/GLM-5.1",
        "https://huggingface.co/zai-org/GLM-5.1-FP8",
        "https://z.ai/blog/glm-5.1",
        "https://docs.z.ai/guides/llm/glm-5.1",
        "https://arxiv.org/abs/2602.15763",
        "https://github.com/zai-org/GLM-5"
      ],
      "page_url": "https://nlp.com.ai/models/glm-51-base-fp8/",
      "record_updated_at": "2026-10-11T08:37:45.556211+00:00"
    },
    {
      "slug": "gemma-4-e2b-e4b-26b-a4b-31b",
      "source_file": "2026-04-02-gemma-4.json",
      "release_date": "2026-04-02",
      "date": "April 2, 2026",
      "name": "Gemma 4 (E2B, E4B, 26B-A4B, 31B)",
      "developer": "Google DeepMind (USA)",
      "parameters": "E2B (~5.1B total), E4B (~8B total), 26B-A4B (25.2B total / 3.8B active), 31B dense (30.7B)",
      "context_window": "128K (E2B/E4B), 256K (26B-A4B/31B)",
      "license": "Apache-2.0",
      "key_features": "Open Gemma 4 family release positioned by Google as its most capable open models to date, focused on advanced reasoning and agentic workflows with high intelligence-per-parameter. Official launch materials present four sizes (E2B, E4B, 26B-A4B MoE, 31B dense) and emphasize broad deployment from edge/mobile to workstation/server settings, with configurable thinking mode, native function calling, structured JSON outputs, and native system prompt support. Model-card documentation highlights multimodal capabilities (text/image across all models, plus audio and video support on E2B/E4B), multilingual coverage (140+ pretraining languages), and benchmark gains across reasoning, coding, and long-context tasks versus prior Gemma generations. Release messaging also stresses a commercially permissive Apache 2.0 license and day-one ecosystem support across major inference and fine-tuning stacks.",
      "paper_url": "https://arxiv.org/abs/2607.02770",
      "links": [
        "https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/",
        "https://huggingface.co/collections/google/gemma-4",
        "https://huggingface.co/google/gemma-4-31B-it",
        "https://arxiv.org/abs/2607.02770",
        "https://ai.google.dev/gemma/docs/core/model_card_4",
        "https://ai.google.dev/gemma/docs/core",
        "https://ai.google.dev/gemma/docs/gemma_4_license"
      ],
      "page_url": "https://nlp.com.ai/models/gemma-4-e2b-e4b-26b-a4b-31b/",
      "record_updated_at": "2026-10-11T08:37:45.554280+00:00"
    },
    {
      "slug": "trinity-large-thinking",
      "source_file": "2026-04-01-trinity-large-thinking.json",
      "release_date": "2026-04-01",
      "date": "April 1, 2026",
      "name": "Trinity-Large-Thinking",
      "developer": "Arcee AI",
      "parameters": "~398B total / ~13B active per token (Sparse MoE, AFMoE)",
      "context_window": "512K",
      "license": "Apache-2.0",
      "key_features": "Open-weight reasoning-optimized Trinity Large variant built from Trinity-Large-Base and post-trained with extended chain-of-thought and agentic RL for tool-use-heavy workflows. Official launch materials position Trinity-Large-Thinking as a frontier open reasoning model for complex long-horizon agents and multi-turn tool calling, and frame it as the official release that closes gaps left by Trinity-Large-Preview. Release claims emphasize stronger multi-turn tool use, better context coherence, cleaner instruction following, and more stable behavior in long-running agent loops, with native reasoning traces in <think></think> blocks that should be preserved in context across turns. Benchmark and deployment messaging highlights reported agentic performance including #2 on PinchBench and competitive scores across IFBench, GPQA-Diamond, AIME25, MMLU-Pro, and SWE-bench Verified-style evaluations, alongside API economics claims of about $0.90 per million output tokens. Architecture and training notes align with the Trinity Large stack: 256 experts with 4 active per token, 512K extended context, and large-scale training/post-training infrastructure, with deployment guidance for vLLM reasoning/tool-call parsers plus compatibility with OpenClaw and Hermes Agent frameworks.",
      "paper_url": "https://arxiv.org/abs/2602.17004",
      "links": [
        "https://huggingface.co/arcee-ai/Trinity-Large-Thinking",
        "https://www.arcee.ai/blog/trinity-large-thinking",
        "https://arxiv.org/abs/2602.17004",
        "https://huggingface.co/arcee-ai/Trinity-Large-Base",
        "https://chat.arcee.ai/",
        "https://openrouter.ai/arcee-ai/trinity-large-thinking"
      ],
      "page_url": "https://nlp.com.ai/models/trinity-large-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.551883+00:00"
    },
    {
      "slug": "lfm25-350m",
      "source_file": "2026-03-31-lfm2.5-350m.json",
      "release_date": "2026-03-31",
      "date": "March 31, 2026",
      "name": "LFM2.5-350M",
      "developer": "Liquid AI",
      "parameters": "350M (LFM2 hybrid architecture; 16 layers: 10 double-gated LIV conv + 6 GQA)",
      "context_window": "",
      "license": "LFM Open License v1.0",
      "key_features": "Open-weight 350M instruction-tuned LFM2.5 release positioned as a small-model breakthrough for on-device AI. Official release materials claim best-in-class quality at this size, highlighting that LFM2.5-350M rivals much larger models while sustaining fast edge inference (reported 313 tok/s decode on AMD CPU and 188 tok/s on Snapdragon Gen4) with sub-1GB memory deployment targets. The model card presents scaled training from 10T to 28T tokens plus large multi-stage reinforcement learning, 32,768 context length, and multilingual support (EN, AR, ZH, FR, DE, JA, KO, PT, ES). Release guidance emphasizes practical use in data extraction, structured outputs, and tool use, with day-one support across Transformers, vLLM, llama.cpp, MLX, and ONNX formats (via companion checkpoints). Benchmark tables in the release report strong gains versus prior LFM2-350M and several other small-model baselines across IFEval/IFBench/Multi-IF and tool/function-calling-oriented evaluations.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2.5-350M",
        "https://www.liquid.ai/blog/lfm2-5-350m-no-size-left-behind",
        "https://arxiv.org/abs/2511.23404",
        "https://docs.liquid.ai"
      ],
      "page_url": "https://nlp.com.ai/models/lfm25-350m/",
      "record_updated_at": "2026-10-11T08:37:45.549612+00:00"
    },
    {
      "slug": "1-bit-bonsai-8b-gguf-mlx",
      "source_file": "2026-03-31-bonsai-8b.json",
      "release_date": "2026-03-31",
      "date": "March 31, 2026",
      "name": "1-bit Bonsai 8B (GGUF, MLX)",
      "developer": "PrismML",
      "parameters": "8.19B (Qwen3-8B dense base; end-to-end 1-bit deployment format)",
      "context_window": "65,536 tokens",
      "license": "Apache-2.0",
      "key_features": "Open-weight 1-bit Bonsai 8B release positioned as the first commercially viable end-to-end 1-bit LLM for real on-device and edge deployment. Official PrismML materials claim true 1-bit coverage across embeddings, attention projections, MLP projections, and LM head with no higher-precision escape path, reducing deployed parameter memory to about 1.15 GB versus 16.38 GB FP16 (about 14x smaller) while remaining competitive with leading full-precision 8B-class instruct models on its benchmark suite. The release highlights major throughput and efficiency claims across runtimes and hardware, including approximately 6.2x faster token generation on RTX 4090 versus FP16 baseline and around 4x-5x lower energy per generated token on reported measurements. Model-card documentation describes cross-platform deployment paths (llama.cpp GGUF on CUDA/Metal/CPU/mobile plus MLX 1-bit for Apple Silicon), a 65,536-token context window, and an intelligence-density framing that presents strong quality-per-GB positioning for latency- and memory-constrained assistants, edge robotics, and private inference scenarios.",
      "paper_url": "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/1-bit-bonsai-8b-whitepaper.pdf",
      "links": [
        "https://prismml.com/news/bonsai-8b",
        "https://huggingface.co/prism-ml/Bonsai-8B-gguf",
        "https://huggingface.co/prism-ml/Bonsai-8B-mlx-1bit",
        "https://github.com/PrismML-Eng/Bonsai-demo/blob/main/1-bit-bonsai-8b-whitepaper.pdf",
        "https://github.com/PrismML-Eng/Bonsai-demo",
        "https://prismml.com/"
      ],
      "page_url": "https://nlp.com.ai/models/1-bit-bonsai-8b-gguf-mlx/",
      "record_updated_at": "2026-10-11T08:37:45.547236+00:00"
    },
    {
      "slug": "gpt-oss-puzzle-88b",
      "source_file": "2026-03-26-gpt-oss-puzzle-88b.json",
      "release_date": "2026-03-26",
      "date": "March 26, 2026",
      "name": "gpt-oss-puzzle-88B",
      "developer": "NVIDIA",
      "parameters": "88B (MoE-derived optimized architecture)",
      "context_window": "",
      "license": "NVIDIA Open Model License",
      "key_features": "Deployment-optimized reasoning and chat model produced with the Puzzle post-training NAS framework to improve inference efficiency under serving constraints. Official model-card claims report substantial throughput gains in both long-context and short-context H100 deployments, while preserving or slightly improving reasoning accuracy across effort settings. The release describes heterogeneous MoE expert pruning, selective window attention for KV-cache efficiency, and RoPE/YaRN scaling adjustments to support stable long-context behavior (128K). Post-architecture optimization includes knowledge distillation and RL refinement with effort-controlled reasoning modes (low/medium/high) for configurable cost-quality trade-offs. Card notes this checkpoint is intended for production-grade, cost-efficient reasoning workloads and includes open deployment guidance and dataset disclosure sections.",
      "paper_url": "https://arxiv.org/abs/2602.11937",
      "links": [
        "https://huggingface.co/nvidia/gpt-oss-puzzle-88B",
        "https://arxiv.org/abs/2411.19146",
        "https://arxiv.org/abs/2602.11937",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/gpt-oss-puzzle-88b/",
      "record_updated_at": "2026-10-11T08:37:45.545010+00:00"
    },
    {
      "slug": "davinci-llm-3b",
      "source_file": "2026-03-26-davinci-llm-3b.json",
      "release_date": "2026-03-26",
      "date": "March 26, 2026",
      "name": "daVinci-LLM-3B",
      "developer": "SII-GAIR-NLP",
      "parameters": "~3.4B (Qwen2-family decoder-only Transformer)",
      "context_window": "",
      "license": "Apache-2.0",
      "key_features": "Open-weight 3B base LLM release positioned as a transparency-first pretraining project. Official model-card claims emphasize full reproducibility artifacts beyond final weights, including training trajectories, intermediate checkpoints, data-processing decisions, and 200+ ablation studies across data quality, mixture design, training dynamics, and evaluation validity. Release materials describe a two-stage ~8T-token curriculum (6T broad web-scale pretraining + 2T structured QA/reasoning data) and claim an overall 51.72 average across 19 benchmarks, reported as approaching or matching some 7B-class baselines such as OLMo-3 7B. The card also frames the model as suitable for research on pretraining science and as a general-purpose base for language, reasoning, and code tasks, while explicitly noting it is a base model that is not instruction- or safety-aligned.",
      "paper_url": "https://arxiv.org/abs/2603.27164",
      "links": [
        "https://huggingface.co/SII-GAIR-NLP/davinci-llm-model"
      ],
      "page_url": "https://nlp.com.ai/models/davinci-llm-3b/",
      "record_updated_at": "2026-10-11T08:37:45.542712+00:00"
    },
    {
      "slug": "longcat-next",
      "source_file": "2026-03-25-longcat-next.json",
      "release_date": "2026-03-25",
      "date": "March 25, 2026",
      "name": "LongCat-Next",
      "developer": "Meituan LongCat",
      "parameters": "A3B (MoE, native multimodal)",
      "context_window": "",
      "license": "MIT",
      "key_features": "Open-source native multimodal autoregressive model that unifies text, vision, and audio in a single discrete token framework under one training objective. Official release materials describe LongCat-Next as an industrial-strength multimodal foundation model built on a LongCat-Flash-Lite MoE backbone (A3B), with support for understanding and generation across modalities plus tool-calling workflows. The project emphasizes a unified discrete representation strategy (including DiNA-style native autoregression and discrete tokenizer/detokenizer interfaces), and reports strong broad multimodal benchmark competitiveness in release documentation. The model card and repo provide end-to-end usage examples spanning text chat, visual understanding/generation, audio tasks, and structured function/tool interactions, and the release includes open model weights together with tokenizer components to support further community research and development.",
      "paper_url": "https://github.com/meituan-longcat/LongCat-Next/blob/main/tech_report.pdf",
      "links": [
        "https://huggingface.co/meituan-longcat/LongCat-Next",
        "https://longcat.chat/longcat-next/intro",
        "https://github.com/meituan-longcat/LongCat-Next",
        "https://github.com/meituan-longcat/LongCat-Next/blob/main/tech_report.pdf",
        "https://longcat.chat/longcat-next"
      ],
      "page_url": "https://nlp.com.ai/models/longcat-next/",
      "record_updated_at": "2026-10-11T08:37:45.540727+00:00"
    },
    {
      "slug": "holo3-35b-a3b",
      "source_file": "2026-03-23-holo3-35b-a3b.json",
      "release_date": "2026-03-23",
      "date": "March 23, 2026",
      "name": "Holo3-35B-A3B",
      "developer": "H Company (France)",
      "parameters": "35B total / 3B active (Sparse MoE)",
      "context_window": "262,144 tokens",
      "license": "Apache-2.0",
      "key_features": "Open-weight Holo3 release positioned as a foundational multimodal model for navigation and computer-use agents across web, desktop, and mobile interfaces. Official materials frame Holo3 as part of H Company's Autonomous Enterprise roadmap, emphasizing an agentic learning flywheel (synthetic navigation data, out-of-domain augmentation, and curated RL) plus synthetic enterprise environments for long-horizon workflow training. The Holo3 launch page reports a family-level 78.85% score on OSWorld-Verified for Holo3-122B-A10B, while the 35B-A3B model card reports 77.8% on OSWorld-Verified with 3B active parameters and positions this checkpoint for cost-efficient enterprise deployment. Release documentation also highlights strong UI localization and grounding performance (including ScreenSpot-Pro and OSWorld-G) and enterprise workflow evaluation via a dedicated multi-step H Corporate Benchmark.",
      "paper_url": "",
      "links": [
        "https://hcompany.ai/holo3",
        "https://huggingface.co/Hcompany/Holo3-35B-A3B",
        "https://huggingface.co/Hcompany/Holo3-35B-A3B/raw/main/README.md",
        "https://huggingface.co/Hcompany/Holo3-35B-A3B/raw/main/config.json"
      ],
      "page_url": "https://nlp.com.ai/models/holo3-35b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.538745+00:00"
    },
    {
      "slug": "nemotron-cascade-2-30b-a3b",
      "source_file": "2026-03-19-nemotron-cascade-2-30b-a3b.json",
      "release_date": "2026-03-19",
      "date": "March 19, 2026",
      "name": "Nemotron-Cascade-2-30B-A3B",
      "developer": "NVIDIA",
      "parameters": "30B / 3B active (MoE)",
      "context_window": "",
      "license": "NVIDIA Open Model License",
      "key_features": "Open 30B MoE model with 3B activated parameters designed for strong reasoning and agentic capabilities in a text-generation interface. Official release notes describe it as post-trained from NVIDIA-Nemotron-3-Nano-30B-A3B-Base and operating in both thinking and instruct (non-reasoning) modes. The release highlights strong competition-level results on math and code reasoning tasks (including IMO/IOI-oriented evaluations and LiveCodeBench-family metrics), alongside broad coverage across STEM, alignment, and long-context evaluations. The model card links openly released supervision artifacts for reproducibility, including SFT and RL datasets, and provides benchmark comparisons against related open models in the same parameter class.",
      "paper_url": "https://arxiv.org/abs/2603.19220",
      "links": [
        "https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B",
        "https://huggingface.co/collections/nvidia/nemotron-cascade-2",
        "https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-SFT-Data",
        "https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-RL-data",
        "https://arxiv.org/abs/2603.19220",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-cascade-2-30b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.536620+00:00"
    },
    {
      "slug": "nemotron-3-nano-4b-bf16-fp8-gguf",
      "source_file": "2026-03-16-nemotron-3-nano-4b.json",
      "release_date": "2026-03-16",
      "date": "March 16, 2026",
      "name": "Nemotron 3 Nano 4B (BF16, FP8, GGUF)",
      "developer": "NVIDIA",
      "parameters": "4B (dense/hybrid SLM)",
      "context_window": "",
      "license": "NVIDIA Nemotron Open Model License",
      "key_features": "Open-weight small language model release in the Nemotron 3 family with BF16, FP8, and GGUF variants for flexible deployment. Official model materials describe Nano 4B as a unified reasoning/non-reasoning model with controllable reasoning behavior (can emit intermediate reasoning traces or produce direct final answers), targeting local assistants, coding support, and edge or application-level automation in English. The model is described as compressed from NVIDIA-Nemotron-Nano-9B-v2 using the Nemotron Elastic framework, with a hybrid architecture primarily composed of Mamba-2 and MLP layers plus a small number of attention layers. Card benchmarks include reasoning-off and reasoning-on evaluations across tool-use, instruction-following, reasoning, and long-context tasks, with published references to BFCL/IFEval/IFBench/Tau2 and AIME25/MATH500/GPQA/LCB style evaluations. Release assets include open model checkpoints plus linked pre-training and post-training dataset collections in the Nemotron ecosystem.",
      "paper_url": "https://huggingface.co/blog/nvidia/nemotron-3-nano-4b",
      "links": [
        "https://huggingface.co/blog/nvidia/nemotron-3-nano-4b",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF",
        "https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets",
        "https://huggingface.co/collections/nvidia/nemotron-post-training-v3",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-3-nano-4b-bf16-fp8-gguf/",
      "record_updated_at": "2026-10-11T08:37:45.534687+00:00"
    },
    {
      "slug": "chroma-context-1",
      "source_file": "2026-03-12-chroma-context-1.json",
      "release_date": "2026-03-12",
      "date": "March 12, 2026",
      "name": "Chroma Context-1",
      "developer": "Chroma",
      "parameters": "20B (MoE, agentic search model)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Open-weight agentic search model trained to retrieve supporting documents for complex multi-hop queries as a retrieval-focused subagent. Official materials describe iterative query decomposition, parallel tool calling, and self-editing context management to maintain retrieval quality over long-horizon search within bounded context budgets. The model is based on gpt-oss-20b and trained with staged SFT plus RL (CISPO), with reported strengths across web/legal/finance retrieval settings and benchmark references including BrowseComp-Plus, SealQA, FRAMES, and HLE. The model card notes an important deployment caveat: the full agent harness used for tool execution, budgeting, pruning, and deduplication is not yet public, so standalone usage may not reproduce reported results.",
      "paper_url": "https://trychroma.com/research/context-1",
      "links": [
        "https://huggingface.co/chromadb/context-1",
        "https://trychroma.com/research/context-1"
      ],
      "page_url": "https://nlp.com.ai/models/chroma-context-1/",
      "record_updated_at": "2026-10-11T08:37:45.532440+00:00"
    },
    {
      "slug": "nemotron-3-super-120b-a12b-base-bf16-fp8-nvfp4",
      "source_file": "2026-03-11-nemotron-3-super.json",
      "release_date": "2026-03-11",
      "date": "March 11, 2026",
      "name": "Nemotron 3 Super 120B-A12B (Base, BF16, FP8, NVFP4)",
      "developer": "NVIDIA",
      "parameters": "120B / 12B active (LatentMoE)",
      "context_window": "",
      "license": "NVIDIA Nemotron Open Model License",
      "key_features": "Open Nemotron 3 Super release family with base and post-trained checkpoints plus optimized precision variants (BF16, FP8, NVFP4); hybrid LatentMoE architecture combining interleaved Mamba-2, MoE, and attention layers with Multi-Token Prediction (MTP); up to 1M-token context window and reasoning mode controls; release positioning emphasizes agentic workflows, tool use, long-context reasoning, and high-throughput enterprise workloads, alongside open dataset collections for pre-training and post-training.",
      "paper_url": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4",
        "https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets",
        "https://huggingface.co/collections/nvidia/nemotron-post-training-v3",
        "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf",
        "https://developer.nvidia.com/nemotron"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-3-super-120b-a12b-base-bf16-fp8-nvfp4/",
      "record_updated_at": "2026-10-11T08:37:45.530417+00:00"
    },
    {
      "slug": "mirothinker-17-and-mirothinker-17-mini",
      "source_file": "2026-03-09-mirothinker-1.7.json",
      "release_date": "2026-03-09",
      "date": "March 9, 2026",
      "name": "MiroThinker-1.7 and MiroThinker-1.7-mini",
      "developer": "MiroMind AI",
      "parameters": "235B and 30B (dense, text-generation)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Open-weight MiroThinker 1.7 release with two checkpoints (1.7 and 1.7-mini) for long-chain agentic reasoning tasks; model card positions the family for deep research workflows with long-horizon reasoning and multi-step tool use; supports up to 256K context and up to 300 tool calls per task, with deployment guidance for SGLang/vLLM and tool-oriented prompting workflows.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/miromind-ai/MiroThinker-1.7",
        "https://huggingface.co/miromind-ai/MiroThinker-1.7-mini",
        "https://huggingface.co/collections/miromind-ai/mirothinker-17",
        "https://github.com/MiroMindAI/MiroThinker"
      ],
      "page_url": "https://nlp.com.ai/models/mirothinker-17-and-mirothinker-17-mini/",
      "record_updated_at": "2026-10-11T08:37:45.528026+00:00"
    },
    {
      "slug": "covenant-72b",
      "source_file": "2026-03-09-covenant-72b.json",
      "release_date": "2026-03-09",
      "date": "March 9, 2026",
      "name": "Covenant-72B",
      "developer": "1Covenant",
      "parameters": "72B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Large open base language model trained from scratch on 1.1T English tokens with a permissionless collaborative-training setup; model card describes 20+ globally distributed participants coordinated over Bittensor and SparseLoCo communication-efficient optimization; LLaMA-style architecture with 80 layers, 64 attention heads (8 KV heads), and 262,144 vocabulary; companion instruction-tuned checkpoint (Covenant-72B-Chat) is released alongside the base model family.",
      "paper_url": "https://arxiv.org/abs/2603.08163",
      "links": [
        "https://huggingface.co/1Covenant/Covenant-72B",
        "https://huggingface.co/1Covenant/Covenant-72B-Chat",
        "https://arxiv.org/abs/2603.08163"
      ],
      "page_url": "https://nlp.com.ai/models/covenant-72b/",
      "record_updated_at": "2026-10-11T08:37:45.525893+00:00"
    },
    {
      "slug": "sarvam-30b-and-105b",
      "source_file": "2026-03-06-sarvam-30b-105b.json",
      "release_date": "2026-03-06",
      "date": "March 6, 2026",
      "name": "Sarvam 30B and 105B",
      "developer": "Sarvam AI (India)",
      "parameters": "30B / 2.4B active (MoE), 105B / 10.3B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Joint open-source release of two MoE reasoning models: Sarvam-30B and Sarvam-105B; both are positioned for multilingual and Indian-language performance with strong coding/reasoning capability. Sarvam-30B uses an efficiency-oriented 19-layer MoE design with top-6 routing over 128 experts, while Sarvam-105B scales to higher capability with top-8 routing over 128 experts, MLA-style attention, and 128K context support via YaRN-style scaling.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/sarvamai/sarvam-30b",
        "https://huggingface.co/sarvamai/sarvam-105b",
        "https://www.sarvam.ai/blogs/sarvam-30b-105b"
      ],
      "page_url": "https://nlp.com.ai/models/sarvam-30b-and-105b/",
      "record_updated_at": "2026-10-11T08:37:45.523857+00:00"
    },
    {
      "slug": "yuan30-ultra",
      "source_file": "2026-03-04-yuan3.0-ultra.json",
      "release_date": "2026-03-04",
      "date": "March 4, 2026",
      "name": "Yuan3.0 Ultra",
      "developer": "YuanLabAI",
      "parameters": "1010B / 68.8B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Trillion-scale multimodal foundation model for enterprise scenarios; unified architecture combining vision encoder, language backbone, and multimodal alignment module; MoE language backbone with 103 Transformer layers; pretrained from scratch at 1515B parameters and reduced to 1010B via Layer-Adaptive Expert Pruning (LAEP), reporting 49% pretraining-efficiency improvement; activates 68.8B parameters per token; incorporates Localized Filtering-based Attention (LFA) and upgraded Reflection Inhibition Reward Mechanism (RIRM); official README reports strong enterprise-oriented benchmark performance across RAG, document/table understanding, coding, and multimodal tasks.",
      "paper_url": "https://github.com/Yuan-lab-LLM/Yuan3.0-Ultra/blob/main/Docs/Yuan3.0_Ultra%20Paper.pdf",
      "links": [
        "https://huggingface.co/YuanLabAI/Yuan3.0-Ultra/tree/main",
        "https://github.com/Yuan-lab-LLM/Yuan3.0-Ultra",
        "https://github.com/Yuan-lab-LLM/Yuan3.0-Ultra/blob/main/Docs/Yuan3.0_Ultra%20Paper.pdf",
        "https://x.com/YuanAI_Lab/status/2029204213180580229"
      ],
      "page_url": "https://nlp.com.ai/models/yuan30-ultra/",
      "record_updated_at": "2026-10-11T08:37:45.521796+00:00"
    },
    {
      "slug": "phi-4-reasoning-vision-15b",
      "source_file": "2026-03-04-phi-4-reasoning-vision-15b.json",
      "release_date": "2026-03-04",
      "date": "March 4, 2026",
      "name": "Phi-4-Reasoning-Vision-15B",
      "developer": "Microsoft",
      "parameters": "15B (multimodal reasoning)",
      "context_window": "",
      "license": "MIT",
      "key_features": "Open-weight multimodal reasoning model built on Phi-4-Reasoning with a SigLIP-2 vision encoder in a mid-fusion architecture; supports image+text input with text output, dynamic image resolution, and up to 3,600 visual tokens for detailed visual understanding; designed for both extended reasoning mode and direct non-reasoning mode, with model-card emphasis on OCR, GUI grounding, computer-use, and math/science reasoning tasks; reported context length is 16,384 tokens with strong benchmark results across AI2D, ChartQA, MathVista, MMMU, OCRBench, and ScreenSpot-v2.",
      "paper_url": "https://aka.ms/Phi-4-reasoning-vision-15B-TR",
      "links": [
        "https://huggingface.co/microsoft/Phi-4-reasoning-vision-15B",
        "https://www.microsoft.com/en-us/research/blog/phi-4-reasoning-vision-and-the-lessons-of-training-a-multimodal-reasoning-model/",
        "https://github.com/microsoft/phi-4-reasoning-vision-15B",
        "https://aka.ms/Phi-4-reasoning-vision-15B-TR"
      ],
      "page_url": "https://nlp.com.ai/models/phi-4-reasoning-vision-15b/",
      "record_updated_at": "2026-10-11T08:37:45.519488+00:00"
    },
    {
      "slug": "step-35-flash-base",
      "source_file": "2026-03-02-step-3.5-flash-base.json",
      "release_date": "2026-03-02",
      "date": "March 2, 2026",
      "name": "Step 3.5 Flash Base",
      "developer": "StepFun",
      "parameters": "196.81B / ~11B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Open foundation checkpoint in the Step 3.5 Flash release pipeline, with companion Base-Midtrain weights released for code, agentic, and long-context workflows; sparse MoE transformer with 45 layers, 256K context window, and 128,896-token vocabulary; routing uses 288 routed experts plus 1 shared expert with Top-8 activation per token; MTP-based generation acceleration and 3:1 sliding-window/full-attention design for efficient long-context inference; StepFun also released the SteptronOss training code and announced SFT data as coming soon.",
      "paper_url": "https://arxiv.org/abs/2602.10604",
      "links": [
        "https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base",
        "https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base-Midtrain",
        "https://x.com/StepFun_ai/status/2028551435290554450",
        "https://github.com/stepfun-ai/SteptronOss",
        "https://arxiv.org/abs/2602.10604"
      ],
      "page_url": "https://nlp.com.ai/models/step-35-flash-base/",
      "record_updated_at": "2026-10-11T08:37:45.517652+00:00"
    },
    {
      "slug": "qwen-35-small-series-qwen35-9b-qwen35-4b-qwen35-2b-qwen35-08b",
      "source_file": "2026-03-02-qwen3.5-small.json",
      "release_date": "2026-03-02",
      "date": "March 2, 2026",
      "name": "Qwen 3.5 Small Series (Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B, Qwen3.5-0.8B)",
      "developer": "Alibaba (Qwen Team)",
      "parameters": "9B, 4B, 2B, 0.8B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Official small-size Qwen 3.5 release set in the Qwen3.5 collection, including Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B, and Qwen3.5-0.8B plus matching Base variants; model cards list image-text-to-text post-trained checkpoints in Hugging Face Transformers format; compatible with Hugging Face Transformers, vLLM, SGLang, and KTransformers; context length is 262,144 tokens with extension guidance up to around 1,010,000 tokens; Qwen3.5 family highlights include unified vision-language training, hybrid Gated DeltaNet plus sparse-MoE design, and large-scale RL generalization.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/collections/Qwen/qwen35",
        "https://huggingface.co/Qwen/Qwen3.5-9B",
        "https://huggingface.co/Qwen/Qwen3.5-4B",
        "https://huggingface.co/Qwen/Qwen3.5-2B",
        "https://huggingface.co/Qwen/Qwen3.5-0.8B",
        "https://qwen.ai/blog?id=qwen3.5"
      ],
      "page_url": "https://nlp.com.ai/models/qwen-35-small-series-qwen35-9b-qwen35-4b-qwen35-2b-qwen35-08b/",
      "record_updated_at": "2026-10-11T08:37:45.516061+00:00"
    },
    {
      "slug": "qwen-35-medium-series-qwen35-35b-a3b-qwen35-122b-a10b-qwen35-27b",
      "source_file": "2026-02-25-qwen3.5-medium.json",
      "release_date": "2026-02-25",
      "date": "February 25, 2026",
      "name": "Qwen 3.5 Medium Series (Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, Qwen3.5-27B)",
      "developer": "Alibaba (Qwen Team)",
      "parameters": "35B / 3B (MoE), 122B / 10B (MoE), 27B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Official mid-tier Qwen 3.5 open-weight lineup focused on balancing capability, speed, and cost; includes Qwen3.5-35B-A3B and Qwen3.5-122B-A10B MoE models plus Qwen3.5-27B dense model; architecture stack combines Gated DeltaNet sequence modeling with QwenMoE/QwenFFN and native multimodal support (image, video, text); context support is 262K natively and can be extended to around 1M tokens via YaRN in many deployments; official release notes position 35B-A3B as outperforming earlier larger Qwen3/Qwen3-VL variants on key benchmarks while 122B-A10B and 27B narrow quality gaps at lower serving cost",
      "paper_url": "",
      "links": [
        "https://huggingface.co/collections/Qwen/qwen35-687aabaf583ee6d4464f1153",
        "https://huggingface.co/Qwen/Qwen3.5-35B-A3B",
        "https://huggingface.co/Qwen/Qwen3.5-122B-A10B",
        "https://huggingface.co/Qwen/Qwen3.5-27B",
        "https://qwen.ai/blog?id=qwen3.5",
        "https://www.linkedin.com/posts/qwenlm_qwen35-medium-is-here-activity-7374344619372711936-zPna",
        "https://qwen.ai"
      ],
      "page_url": "https://nlp.com.ai/models/qwen-35-medium-series-qwen35-35b-a3b-qwen35-122b-a10b-qwen35-27b/",
      "record_updated_at": "2026-10-11T08:37:45.514472+00:00"
    },
    {
      "slug": "lfm2-24b-a2b",
      "source_file": "2026-02-24-lfm2-24b-a2b.json",
      "release_date": "2026-02-24",
      "date": "February 24, 2026",
      "name": "LFM2-24B-A2B",
      "developer": "Liquid AI",
      "parameters": "24B / 2.3B active (MoE)",
      "context_window": "",
      "license": "LFM Open License v1.0",
      "key_features": "Largest LFM2 model at release; sparse Mixture-of-Experts architecture with 24B total parameters and ~2.3B active per token; hybrid backbone combining gated short convolutions with grouped-query attention (40 layers total: 30 conv + 10 attention) and top-4 routing across 64 experts per MoE block; designed to fit in ~32GB RAM for deployment on consumer laptops/desktops and edge hardware; 32,768 context window; day-zero inference support across llama.cpp, vLLM, and SGLang with GGUF quantization options; benchmark reports show log-linear quality scaling across GPQA Diamond, MMLU-Pro, IFEval, IFBench, GSM8K, and MATH-500 as the LFM2 family scales up",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2-24B-A2B",
        "https://www.liquid.ai/blog/lfm2-24b-a2b",
        "https://www.liquid.ai/blog/lfm2-24b-a2b-from-cloud-to-ai-pc",
        "https://docs.liquid.ai"
      ],
      "page_url": "https://nlp.com.ai/models/lfm2-24b-a2b/",
      "record_updated_at": "2026-10-11T08:37:45.512129+00:00"
    },
    {
      "slug": "tiny-aya",
      "source_file": "2026-02-17-tiny-aya.json",
      "release_date": "2026-02-17",
      "date": "February 17, 2026",
      "name": "Tiny Aya",
      "developer": "Cohere Labs (Canada)",
      "parameters": "3.35B (dense)",
      "context_window": "",
      "license": "CC-BY-NC (non-commercial; contact Cohere for commercial use)",
      "key_features": "Massively multilingual small language model supporting 70+ languages including many low-resource ones; optimized for efficient local/on-device deployment including mobile; 8K context with 8K output length; trained on single cluster of 64 H100 GPUs; state-of-the-art translation quality; regional variants: TinyAya-Global (instruction-tuned), TinyAya-Earth (Africa), TinyAya-Fire (South Asia), TinyAya-Water (Asia-Pacific/Europe); supports Indic languages (Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, Marathi), East/Southeast Asian, African, European, Middle Eastern language families; launched at India AI Summit",
      "paper_url": "",
      "links": [
        "https://huggingface.co/CohereLabs/tiny-aya-global",
        "https://cohere.com/blog/cohere-labs-tiny-aya"
      ],
      "page_url": "https://nlp.com.ai/models/tiny-aya/",
      "record_updated_at": "2026-10-11T08:37:45.510443+00:00"
    },
    {
      "slug": "qwen35-397b-a17b",
      "source_file": "2026-02-16-qwen3.5-397b-a17b.json",
      "release_date": "2026-02-16",
      "date": "February 16, 2026",
      "name": "Qwen3.5-397B-A17B",
      "developer": "Alibaba (Qwen Team)",
      "parameters": "397B / 17B (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Next-generation flagship model for the agentic AI era; native multimodal (image-text-to-text); 60% cheaper and 8x more efficient at processing large workloads vs Qwen3; 256K context window; visual agentic capabilities for autonomous actions across mobile and desktop apps; outperforms GPT-5.2, Claude Opus 4.5, Gemini 3 Pro on benchmarks; released on eve of Lunar New Year amid major Chinese AI release wave",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Qwen/Qwen3.5-397B-A17B",
        "https://qwen.ai/blog?id=qwen3.5",
        "https://qwen.ai"
      ],
      "page_url": "https://nlp.com.ai/models/qwen35-397b-a17b/",
      "record_updated_at": "2026-10-11T08:37:45.508251+00:00"
    },
    {
      "slug": "ling-25-1t",
      "source_file": "2026-02-16-ling-2.5-1t.json",
      "release_date": "2026-02-16",
      "date": "February 16, 2026",
      "name": "Ling-2.5-1T",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "1T / 63B (MoE)",
      "context_window": "",
      "license": "MIT",
      "key_features": "Flagship instant model in Ling family with hybrid linear attention architecture; transforms GQA layers into Lightning Linear Attention for enhanced throughput + converts remaining GQA to MLA with QK Norm and Partial RoPE for KV cache compression; 256K base context extending to 1M via YaRN; pre-trained on 29T tokens (up from 20T in Ling 2.0); reasoning approaches frontier thinking models at ~4x fewer output tokens; trained with Agentic RL in large-scale high-fidelity interactive environments; compatible with Claude Code, OpenCode, OpenClaw; leading open-source on BFCL-V4 (general tool-calling); competitive with DeepSeek V3.2, Kimi K2.5, GPT 5.2 in complex reasoning and instruction-following; high-throughput decoding for ultra-long context",
      "paper_url": "",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-2.5-1T",
        "https://modelscope.cn/models/inclusionAI/Ling-2.5-1T",
        "https://github.com/inclusionAI/Ling"
      ],
      "page_url": "https://nlp.com.ai/models/ling-25-1t/",
      "record_updated_at": "2026-10-11T08:37:45.506476+00:00"
    },
    {
      "slug": "ring-25-1t",
      "source_file": "2026-02-12-ring-2.5-1t.json",
      "release_date": "2026-02-12",
      "date": "February 12, 2026",
      "name": "Ring-2.5-1T",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "1T / 63B (MoE)",
      "context_window": "",
      "license": "MIT",
      "key_features": "World's first open-source trillion-parameter reasoning model based on hybrid linear attention architecture; 1:7 MLA + Lightning Linear Attention ratio reduces memory access overhead by 10x and increases generation throughput by 3x for sequences exceeding 32K tokens; 128K context (expandable to 256K via YaRN); enhanced deep reasoning with dense rewards for rigor feedback during RLVR; large-scale fully-async agentic RL training for long-horizon task execution; IMO 2025 gold medal level (35/42 points); CMO 2025 (105/126 points, exceeding gold threshold of 78); SOTA on Gaia2-search among open-source models; compatible with Claude Code and OpenClaw frameworks; supports agentic programming, multi-step tool calling, cross-application tool collaboration",
      "paper_url": "",
      "links": [
        "https://huggingface.co/inclusionAI/Ring-2.5-1T",
        "https://modelscope.cn/models/inclusionAI/Ring-2.5-1T",
        "https://github.com/inclusionAI/Ring-V2.5"
      ],
      "page_url": "https://nlp.com.ai/models/ring-25-1t/",
      "record_updated_at": "2026-10-11T08:37:45.504215+00:00"
    },
    {
      "slug": "minimax-m25",
      "source_file": "2026-02-12-minimax-m2.5.json",
      "release_date": "2026-02-12",
      "date": "February 12, 2026",
      "name": "MiniMax-M2.5",
      "developer": "MiniMax (China)",
      "parameters": "230B / 10B (MoE)",
      "context_window": "",
      "license": "MIT (modified, requires 'MiniMax M2.5' attribution on commercial product UI)",
      "key_features": "Frontier model designed for real-world productivity; extensively trained with RL in 200K+ complex real-world environments; SOTA in coding, agentic tool use, search, and office work; SWE-Bench Verified (80.2%), Multi-SWE-Bench (51.3%), BrowseComp (76.3%); two variants: M2.5 and M2.5-Lightning (2x faster at 100 tok/s steady throughput); 205K context window; trained on 10+ programming languages (Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby); cost efficiency: $1 for 1 hour continuous at 100 tok/s, 1/10th to 1/20th cost of Opus/Gemini 3 Pro/GPT-5; pricing: $0.30/1M input, $1.20/1M output",
      "paper_url": "",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M2.5",
        "https://www.minimax.io/news/minimax-m25",
        "https://artificialanalysis.ai/models/minimax-m2-5"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m25/",
      "record_updated_at": "2026-10-11T08:37:45.501960+00:00"
    },
    {
      "slug": "nanbeige41-3b",
      "source_file": "2026-02-11-nanbeige4.1-3b.json",
      "release_date": "2026-02-11",
      "date": "February 11, 2026",
      "name": "Nanbeige4.1-3B",
      "developer": "Nanbeige (China)",
      "parameters": "3B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Compact reasoning model built on Nanbeige4-3B-Base with further SFT and RL post-training; pretrained on 23T high-quality tokens and finetuned on 30M+ diverse instructions; 64K context via RoPE with ABF technique; strong multi-step reasoning rivaling 30B models like Qwen3-32B; first general small model to natively support deep-search tasks with 500+ rounds of tool invocations; Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler; Dual Preference Distillation (DPD) method; strong benchmarks: AIME 2026 I (87.4%), LiveCodeBench-Pro Easy (81.4%), Arena-Hard-v2 (73.2%), IMO-Answer-Bench; outperforms Qwen3-4B, Qwen3-8B, Qwen3-30B-A3B, Qwen3-32B on Arena-Hard-v2 and Multi-Challenge; supports chat, tool use, and deep-search scenarios",
      "paper_url": "https://arxiv.org/abs/2512.06266",
      "links": [
        "https://huggingface.co/Nanbeige/Nanbeige4.1-3B",
        "https://arxiv.org/abs/2512.06266",
        "https://x.com/nanbeige/status/2021471995662303518"
      ],
      "page_url": "https://nlp.com.ai/models/nanbeige41-3b/",
      "record_updated_at": "2026-10-11T08:37:45.499545+00:00"
    },
    {
      "slug": "minicpm-sala",
      "source_file": "2026-02-11-minicpm-sala.json",
      "release_date": "2026-02-11",
      "date": "February 11, 2026",
      "name": "MiniCPM-SALA",
      "developer": "OpenBMB / ModelBest (China)",
      "parameters": "9B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "First 9B model trained with Sparse-Linear Attention (SALA) hybrid architecture; 75% linear attention (Lightning Attention) for global efficiency + 25% sparse attention (InfLLM-v2) for high-fidelity long context modeling; 1M+ token context via HyPE (Hybrid Positional Embedding); 3.5x inference speed and significantly lower KV-cache overhead vs dense baselines; first model to achieve million-token context inference on consumer GPUs (RTX 5090); maintains 81.6 score at 2048K context despite training on only 520K tokens; NoPE configuration in sparse attention layers for length extrapolation without YaRN; average score 76.53 across standard benchmarks; outperforms Qwen3-8B and Falcon-H1R-7B; strong in knowledge, code, and math",
      "paper_url": "https://github.com/OpenBMB/MiniCPM/blob/main/docs/MiniCPM_SALA.pdf",
      "links": [
        "https://huggingface.co/openbmb/MiniCPM-SALA",
        "https://github.com/OpenBMB/MiniCPM",
        "https://github.com/OpenBMB/MiniCPM/blob/main/docs/MiniCPM_SALA.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/minicpm-sala/",
      "record_updated_at": "2026-10-11T08:37:45.497334+00:00"
    },
    {
      "slug": "ming-flash-omni-20",
      "source_file": "2026-02-11-ming-flash-omni-2.0.json",
      "release_date": "2026-02-11",
      "date": "February 11, 2026",
      "name": "Ming-flash-omni-2.0",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "100B / 6B (MoE)",
      "context_window": "",
      "license": "MIT",
      "key_features": "SOTA open-source omni-MLLM built on Ling-2.0 MoE architecture; supports image, text, video, and audio inputs/outputs; expert-level multimodal cognition with vision-to-knowledge synthesis (plant/animal identification, cultural references, artifact analysis); first industry-wide unified audio generation model producing speech, ambient sound effects, and music on same audio track; ultra-low inference frame rate of 3.1Hz for real-time minute-long audio generation; zero-shot voice cloning with natural language control of voice, speed, intonation, volume, emotion, dialect; Continuous Autoregression with Diffusion Transformer (DiT) head; high-dynamic controllable image generation with native multi-task architecture unifying segmentation, generation, and editing; surpasses Gemini 2.5 Pro in some metrics",
      "paper_url": "https://arxiv.org/abs/2506.09344",
      "links": [
        "https://huggingface.co/inclusionAI/Ming-flash-omni-2.0",
        "https://modelscope.cn/models/inclusionAI/Ming-flash-omni-2.0",
        "https://github.com/inclusionAI/Ming"
      ],
      "page_url": "https://nlp.com.ai/models/ming-flash-omni-20/",
      "record_updated_at": "2026-10-11T08:37:45.495066+00:00"
    },
    {
      "slug": "glm-5",
      "source_file": "2026-02-11-glm-5.json",
      "release_date": "2026-02-11",
      "date": "February 11, 2026",
      "name": "GLM-5",
      "developer": "Zhipu AI / Z.AI (China)",
      "parameters": "744B / 40B (MoE)",
      "context_window": "",
      "license": "MIT",
      "key_features": "Flagship MoE model trained entirely on Huawei Ascend 910 chips with MindSpore framework; 256 experts with 8 activated per token; Dynamically Sparse Attention (DSA) mechanism using DeepSeek Sparse Attention; 200K context window with 131K max output; agentic intelligence mode for long-running tasks; specialized in creative writing, coding, advanced reasoning, and long-context processing; #1 open-source model globally at release; strong benchmarks: SWE-bench Verified (77.8%), Humanity's Last Exam (50.4), τ²-Bench (89.7), BrowseComp (75.9); approaches Claude Opus 4.5 in coding benchmarks; surpasses Gemini 3 Pro on some benchmarks; supports data insights, smart writing, full-stack development; 17-19 tokens/sec throughput; zero dependency on US-manufactured hardware",
      "paper_url": "https://arxiv.org/abs/2602.15763",
      "links": [
        "https://huggingface.co/zai-org/GLM-5",
        "https://chat.z.ai",
        "https://openrouter.ai/z-ai/glm-5",
        "https://z.ai/blog/glm-5",
        "https://arxiv.org/abs/2602.15763",
        "https://github.com/zai-org/GLM-5"
      ],
      "page_url": "https://nlp.com.ai/models/glm-5/",
      "record_updated_at": "2026-10-11T08:37:45.493067+00:00"
    },
    {
      "slug": "llada21-mini",
      "source_file": "2026-02-09-llada2.1-mini.json",
      "release_date": "2026-02-09",
      "date": "February 9, 2026",
      "name": "LLaDA2.1-mini",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "16B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Diffusion language model with token-editing capability allowing self-correction during generation; generates tokens in parallel and refines through diffusion-style process rather than strict autoregressive generation; two operational modes: Speedy Mode (S Mode) for speed with T2T refinement, Quality Mode (Q Mode) for superior benchmark performance; peak inference speed of 1586.93 TPS after quantization; 32K context window; 20 layers, 16 attention heads, 157K vocabulary; strong benchmarks: MMLU-Pro, C-EVAL (knowledge), BIG-Bench Hard, DROP (reasoning), HumanEval+, Spider (coding), CMATH, GSM-Plus (math); overall average score 63.90; requires ~48GB VRAM",
      "paper_url": "https://arxiv.org/abs/2602.08676",
      "links": [
        "https://huggingface.co/inclusionAI/LLaDA2.1-mini",
        "https://modelscope.cn/models/inclusionAI/LLaDA2.1-mini",
        "https://github.com/inclusionAI/LLaDA2.X"
      ],
      "page_url": "https://nlp.com.ai/models/llada21-mini/",
      "record_updated_at": "2026-10-11T08:37:45.491137+00:00"
    },
    {
      "slug": "qwen3-coder-next",
      "source_file": "2026-02-03-qwen3-coder-next.json",
      "release_date": "2026-02-03",
      "date": "February 3, 2026",
      "name": "Qwen3-Coder-Next",
      "developer": "Qwen Team / Alibaba (China)",
      "parameters": "80B total / 3B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Open-weight coding-focused model designed for coding agents and local development with high efficiency; model card states performance comparable to models with 10-20x more active parameters while activating only 3B out of 80B total. Architecture is a 48-layer hybrid stack following 12 × (3 × (Gated DeltaNet -> MoE) -> 1 × (Gated Attention -> MoE)); Gated DeltaNet uses 32 linear attention heads for V and 16 for QK (head dim 128), while Gated Attention uses 16 Q heads and 2 KV heads (head dim 256, RoPE dim 64). MoE configuration uses 512 experts with 10 activated experts plus 1 shared expert per token and expert intermediate dim 512. Native context length is 262,144 tokens for repository-scale code understanding. Release notes emphasize advanced agentic behavior for long-horizon reasoning, complex tool usage, and recovery from execution failures, plus integration across developer tooling and scaffolds (including CLI/IDE environments such as Claude Code, Qwen Code, Qoder, Kilo, Trae, and Cline). The model is non-thinking mode only and does not generate <think></think> blocks. Deployment guidance highlights support in modern serving stacks including SGLang and vLLM, with additional ecosystem support referenced for other runtimes.",
      "paper_url": "https://arxiv.org/abs/2603.00729",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-Coder-Next",
        "https://qwenlm.github.io/blog/qwen3-coder-next/",
        "https://github.com/QwenLM/Qwen3-Coder"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-coder-next/",
      "record_updated_at": "2026-10-11T08:37:45.489060+00:00"
    },
    {
      "slug": "minicpm-o-45",
      "source_file": "2026-02-03-minicpm-o-4.5.json",
      "release_date": "2026-02-03",
      "date": "February 3, 2026",
      "name": "MiniCPM-o 4.5",
      "developer": "OpenBMB / ModelBest (China)",
      "parameters": "9B",
      "context_window": "40,960 tokens",
      "license": "Apache 2.0",
      "key_features": "Official open-weight omnimodal MiniCPM release built end-to-end from SigLip2, Whisper-medium, CosyVoice2, and Qwen3-8B into a single 9B-parameter model. Release materials position MiniCPM-o 4.5 as the strongest model in the MiniCPM-o line, supporting text, image, video, and audio inputs with text and speech outputs, plus both instruct and thinking modes in one checkpoint. The model card highlights full-duplex multimodal live streaming so video/audio input and text/speech output can run concurrently without blocking, along with proactive interaction, bilingual real-time speech conversation in English and Chinese, configurable voices, and zero-shot voice cloning. Official benchmarks claim 77.6 average on an 8-benchmark OpenCompass vision evaluation, text average 82.1 across IFEval-PLS, BBH, CMMLU, MMLU, HumanEval, MBPP, Math500, and GSM8K, strong OmniDocBench document parsing, MMHal-Bench trustworthiness near Gemini 2.5 Flash, and broad multilingual support across 30+ languages. The released ecosystem also emphasizes practical deployment through PyTorch, vLLM, SGLang, llama.cpp, Ollama, AWQ and GGUF quantizations, WebRTC demo tooling, and FlagOS multi-chip adaptations.",
      "paper_url": "https://arxiv.org/abs/2408.01800",
      "links": [
        "https://huggingface.co/openbmb/MiniCPM-o-4_5",
        "https://huggingface.co/openbmb/MiniCPM-o-4_5/blob/main/config.json",
        "https://huggingface.co/openbmb/MiniCPM-o-4_5-awq",
        "https://huggingface.co/openbmb/MiniCPM-o-4_5-gguf",
        "https://github.com/OpenBMB/MiniCPM-o",
        "https://www.openbmb.cn/en/about-us",
        "https://modelbest.cn/en",
        "https://www.linkedin.com/company/modelbest/",
        "https://arxiv.org/abs/2408.01800"
      ],
      "page_url": "https://nlp.com.ai/models/minicpm-o-45/",
      "record_updated_at": "2026-10-11T08:37:45.487159+00:00"
    },
    {
      "slug": "step-35-flash",
      "source_file": "2026-02-02-step-3.5-flash.json",
      "release_date": "2026-02-02",
      "date": "February 2, 2026",
      "name": "Step 3.5 Flash",
      "developer": "StepFun (China)",
      "parameters": "196.81B / ~11B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "StepFun flagship open foundation model positioned for frontier reasoning and agentic capabilities with high efficiency; sparse MoE architecture activates roughly 11B out of 196.81B total parameters per token; designed for AI agents, agentic coding, deep research, and tool orchestration workflows. Model card materials describe MTP-3 generation with typical 100-300 tok/s throughput (peaking around 350 tok/s in single-stream coding scenarios), 256K context via a 3:1 sliding-window/full-attention hybrid, and reasoning-mode control through chat-template configuration. Technical details highlight a 45-layer transformer backbone, 128,896-token vocabulary, and expert routing across 288 routed experts plus 1 shared expert with Top-8 activation per token. Reported benchmark highlights include AIME 2025, HMMT 2025, IMOAnswerBench, SWE-bench Verified (74.4), Terminal-Bench 2.0 (51.0), LiveCodeBench-V6, ResearchRubrics, τ²-Bench, and BrowseComp with context-manager settings. Release positioning emphasizes robust long-horizon agent execution, tool-use reliability, and cost-efficient decoding relative to larger-parameter peers, with deployment support across major open inference stacks and community integrations.",
      "paper_url": "https://arxiv.org/abs/2602.10604",
      "links": [
        "https://github.com/stepfun-ai/Step-3.5-Flash",
        "https://huggingface.co/stepfun-ai/Step-3.5-Flash",
        "https://static.stepfun.com/blog/step-3.5-flash/",
        "https://arxiv.org/abs/2602.10604"
      ],
      "page_url": "https://nlp.com.ai/models/step-35-flash/",
      "record_updated_at": "2026-10-11T08:37:45.484712+00:00"
    },
    {
      "slug": "olmo-hybrid-7b-series-base-think-sft-instruct-sft-instruct-dpo-think",
      "source_file": "2026-01-28-olmo-hybrid-series.json",
      "release_date": "2026-01-28",
      "date": "January 28, 2026",
      "name": "OLMo Hybrid 7B Series (Base, Think-SFT, Instruct-SFT, Instruct-DPO, Think)",
      "developer": "Allen Institute for AI (Ai2)",
      "parameters": "7B (hybrid RNN/Transformer dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Hybrid OLMo-family release with staged checkpoints covering Base, Think-SFT, Instruct-SFT, Instruct-DPO, and Think tracks; architecture alternates 3 gated DeltaNet layers for every 1 attention layer (75% DeltaNet, 25% attention), with 32 layers and 65,536 context length; model-card claims include roughly 2x data efficiency over OLMo 3 and major long-context throughput/memory efficiency gains; release includes training data details and published datasets for pretraining, mid-training, and long-context stages (Dolma3 and Dolmino/Longmino mixes).",
      "paper_url": "https://allenai.org/papers/olmo-hybrid",
      "links": [
        "https://huggingface.co/allenai/Olmo-Hybrid-7B",
        "https://huggingface.co/allenai/Olmo-Hybrid-Think-SFT-7B",
        "https://huggingface.co/allenai/Olmo-Hybrid-Instruct-SFT-7B",
        "https://huggingface.co/allenai/Olmo-Hybrid-Instruct-DPO-7B",
        "https://huggingface.co/datasets/allenai/dolma3_mix-6T",
        "https://huggingface.co/datasets/allenai/dolma3_dolmino_mix-100B-1025",
        "https://huggingface.co/datasets/allenai/dolma3_longmino_mix-50B-1025",
        "https://allenai.org/papers/olmo-hybrid",
        "https://github.com/allenai/OLMo"
      ],
      "page_url": "https://nlp.com.ai/models/olmo-hybrid-7b-series-base-think-sft-instruct-sft-instruct-dpo-think/",
      "record_updated_at": "2026-10-11T08:37:45.482312+00:00"
    },
    {
      "slug": "trinity-large-preview-base-truebase",
      "source_file": "2026-01-27-trinity-large.json",
      "release_date": "2026-01-27",
      "date": "January 27, 2026",
      "name": "Trinity Large (Preview, Base, TrueBase)",
      "developer": "Arcee AI (USA)",
      "parameters": "400B total / 13B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Flagship Trinity family release with three checkpoints from the same run: Trinity-Large-Preview (lightly post-trained, chat-ready), Trinity-Large-Base (full 17T-token pretrained checkpoint), and Trinity-Large-TrueBase (10T-token pre-anneal checkpoint without instruct data). Release materials describe Trinity Large as a sparse MoE model with approximately 13B active parameters per token, using 256 experts with 4 active per token (4-of-256 routing, ~1.56% sparsity), plus a shared expert. Architecture and training notes reference AFMoE design, HSDP plus expert parallelism, and large-scale infrastructure using 2,048 NVIDIA B300 GPUs; official blog text states pretraining completed in 33 days and cites an all-in project cost around $20M. Data curation is attributed to DatologyAI with a 17T-token recipe across phased mixes (including code, STEM, reasoning, and multilingual content), with release messaging noting 14 non-English languages. Context support is presented as 512K native, with preview API operation at 128K under 8-bit quantized serving. Benchmark tables in release materials highlight strong base-model performance including MMLU 87.2 and competitive Math-500 positioning, while Preview is positioned for interactive agentic and coding-oriented deployment across open inference stacks and hosted endpoints.",
      "paper_url": "https://arxiv.org/abs/2602.17004",
      "links": [
        "https://huggingface.co/arcee-ai/Trinity-Large-Preview",
        "https://huggingface.co/arcee-ai/Trinity-Large-Base",
        "https://huggingface.co/arcee-ai/Trinity-Large-TrueBase",
        "https://www.arcee.ai/blog/trinity-large",
        "https://arxiv.org/abs/2602.17004",
        "https://github.com/arcee-ai/trinity-large-tech-report/"
      ],
      "page_url": "https://nlp.com.ai/models/trinity-large-preview-base-truebase/",
      "record_updated_at": "2026-10-11T08:37:45.479905+00:00"
    },
    {
      "slug": "sera-32b",
      "source_file": "2026-01-27-sera-32b.json",
      "release_date": "2026-01-27",
      "date": "January 27, 2026",
      "name": "SERA-32B",
      "developer": "Allen Institute for AI (Ai2) - USA",
      "parameters": "32B (based on Qwen 3-32B)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "First release in Ai2's Open Coding Agents series, positioned as a specialized software-engineering agent rather than a general chat model. Official model-card reporting cites 49.5% ± 1.9% on SWE-bench Verified at 32K context (with related comparisons showing competitiveness against larger open-weight models). Training uses Soft Verified Generation (SVG), a two-rollout synthetic trajectory pipeline where a teacher model edits code, a synthetic PR description is produced, and a second rollout reproduces the change; card claims SVG reaches target quality at substantially lower cost (26x cheaper than RL and 57x cheaper than prior synthetic-data methods) with total generation+training cost around $2,000 (40 GPU-days). The released training set size is 25,000 synthetic coding-agent trajectories. Deployment guidance includes vLLM-based self-hosting and sera-cli integration with Claude Code, with references to Modal and endpoint-based operation. The release family also includes SERA-32B-GA (46.6%), SERA-8B (31.7%), and SERA-8B-GA (31.7%), with teacher-model differences noted in the card. Limitations section flags a known SWE-bench artifact where the model may attempt a nonexistent `submit` tool call in some workflows.",
      "paper_url": "https://allenai.org/papers/opencodingagents",
      "links": [
        "https://huggingface.co/allenai/SERA-32B",
        "https://huggingface.co/collections/allenai/open-coding-agents",
        "https://huggingface.co/datasets/allenai/Sera-4.6-Lite-T2",
        "https://github.com/allenai/SERA",
        "https://github.com/allenai/sera-cli",
        "https://allenai.org/papers/opencodingagents"
      ],
      "page_url": "https://nlp.com.ai/models/sera-32b/",
      "record_updated_at": "2026-10-11T08:37:45.477525+00:00"
    },
    {
      "slug": "davinci-dev-32b-72b",
      "source_file": "2026-01-26-davinci-dev.json",
      "release_date": "2026-01-26",
      "date": "January 26, 2026",
      "name": "daVinci-Dev (32B, 72B)",
      "developer": "GAIR (SII)",
      "parameters": "32B and 72B (plus MT checkpoints)",
      "context_window": "",
      "license": "Qwen license (models); mixed release components for data/pipeline include CC-BY-4.0 and Apache-2.0",
      "key_features": "Open-weight agentic software-engineering model family positioned as agent-native mid-training for repository-level coding workflows (navigate/edit/test loops) rather than single-turn code generation. Official release materials emphasize a scalable training recipe that combines contextually-native PR trajectories and environmentally-native executable trajectories, then applies environment-native SFT. The model card claims state-of-the-art performance among open training recipes with agentic scaffolds at these model sizes, reporting SWE-Bench Verified Pass@1 of 56.1% (32B) and 58.5% (72B), evaluated with SWE-Agent at temperature 0, 128k context, and a 100-step budget; release notes also claim generalization gains on HumanEval/EvalPlus and GPQA/SciBench. The release includes model checkpoints, dataset artifacts, and pipeline components for reproducible agentic training workflows.",
      "paper_url": "https://arxiv.org/abs/2601.18418",
      "links": [
        "https://arxiv.org/abs/2601.18418",
        "https://github.com/GAIR-NLP/daVinci-Dev",
        "https://huggingface.co/GAIR/daVinci-Dev-72B",
        "https://huggingface.co/GAIR/daVinci-Dev-32B",
        "https://huggingface.co/datasets/GAIR/daVinci-Dev"
      ],
      "page_url": "https://nlp.com.ai/models/davinci-dev-32b-72b/",
      "record_updated_at": "2026-10-11T08:37:45.474884+00:00"
    },
    {
      "slug": "mistral-small-4-119b-base-nvfp4-eagle",
      "source_file": "2026-01-23-mistral-small-4.json",
      "release_date": "2026-01-23",
      "date": "January 23, 2026",
      "name": "Mistral Small 4 119B (Base, NVFP4, Eagle)",
      "developer": "Mistral AI",
      "parameters": "119B / 6.5B active (MoE)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Open-weight multimodal MoE model positioned as a unified successor across Mistral's Instruct, Reasoning (Magistral), and Devstral usage patterns in one checkpoint family. Official release materials describe Mistral Small 4 as the first Small-family model to combine fast instruct behavior, on-demand deep reasoning, and agentic coding capabilities in a single system, alongside multimodal text+image input support. Architecture details cite 128 experts with 4 active experts per token, 119B total parameters, and approximately 6B-6.5B activated parameters per token (with release notes also describing around 8B active including embedding/output layers), plus a 256K context window for long-form tasks. The model supports configurable reasoning via per-request `reasoning_effort` controls, allowing fast low-latency operation (`none`) or deeper stepwise reasoning (`high`) depending on task complexity. Official positioning emphasizes strong function-calling and structured-output workflows, multilingual operation, and broad applicability across chat, coding, tool-use, and document/image analysis scenarios. Release materials also provide efficiency-oriented companion checkpoints: an NVFP4 quantized variant for lower-cost inference and an Eagle speculative-decoding head for faster generation. Announcement and card claims include improved serving efficiency relative to prior Small-family baselines, including reported 40% reduction in end-to-end completion time and 3x higher request throughput in optimized setups.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/mistral-small-4",
        "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603",
        "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603-NVFP4",
        "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603-eagle"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-small-4-119b-base-nvfp4-eagle/",
      "record_updated_at": "2026-10-11T08:37:45.472625+00:00"
    },
    {
      "slug": "antangelmed",
      "source_file": "2026-01-07-antangelmed.json",
      "release_date": "2026-01-07",
      "date": "January 7, 2026",
      "name": "AntAngelMed",
      "developer": "Ant Group / Health Information Center of Zhejiang Province (China)",
      "parameters": "100B / 6.1B active (MoE, 1/32 activation ratio)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Large open-source medical language model based on Ling-flash-2.0, jointly developed by healthcare and industry partners for professional medical use cases. Official release materials position the model as top-performing on major medical benchmarks, including #1 among open-source models on HealthBench and leading results on MedBench and MedAIBench, with emphasis on medical knowledge QA, diagnostic reasoning, and safety/ethics performance. Training is described as a three-stage pipeline: continual pretraining on medical corpora, supervised fine-tuning with high-quality medical instructions, and GRPO-based reinforcement learning for robust clinical reasoning behavior. Architecture uses high-efficiency sparse MoE design with 100B total parameters and about 6.1B active parameters (1/32 activation ratio), targeting strong capability at lower inference cost; release text reports over 200 tokens/s on H20 hardware and 128K context support. The project also publishes FP8 quantized variants and reports EAGLE3-based acceleration gains under concurrent inference workloads, with specific throughput improvements on HumanEval, GSM8K, and Math-500 settings. Distribution includes Hugging Face, ModelScope, project GitHub, and a public API demo endpoint.",
      "paper_url": "https://huggingface.co/MedAIBase/AntAngelMed",
      "links": [
        "https://huggingface.co/MedAIBase/AntAngelMed",
        "https://huggingface.co/MedAIBase/AntAngelMed-FP8",
        "https://modelscope.cn/models/MedAIBase/AntAngelMed",
        "https://github.com/MedAIBase/AntAngelMed",
        "https://antangelmed.tbox.cn/",
        "https://arxiv.org/abs/2507.17702",
        "https://arxiv.org/abs/2505.08775",
        "https://arxiv.org/abs/2511.14439"
      ],
      "page_url": "https://nlp.com.ai/models/antangelmed/",
      "record_updated_at": "2026-10-11T08:37:45.470189+00:00"
    },
    {
      "slug": "nouscoder-14b",
      "source_file": "2026-01-06-nouscoder-14b.json",
      "release_date": "2026-01-06",
      "date": "January 6, 2026",
      "name": "NousCoder-14B",
      "developer": "Nous Research (USA)",
      "parameters": "14B (dense)",
      "context_window": "",
      "license": "Apache 2.0",
      "key_features": "Competitive coding-focused model post-trained from Qwen3-14B using reinforcement learning, positioned for olympiad-style and verifiable programming tasks. Official release materials report 67.87% Pass@1 on LiveCodeBench v6, a +7.08 point gain over the stated Qwen3-14B baseline (60.79%). Training is described as using 24k verifiable coding problems on 48 B200 GPUs over approximately four days, with a fully released stack that includes model weights, open RL environment, evaluation harness, and experiment logs. The release details iterative context-length training behavior (including extension from 32k to 40k context during training) and evaluation references at longer windows, plus a pipelined verification setup and parallelization notes intended to support reproducibility. Distribution is provided as open-weight text-generation model artifacts on Hugging Face under Apache 2.0, with accompanying technical write-up and code resources in the official Nous release channels.",
      "paper_url": "https://nousresearch.com/nouscoder-14b-a-competitive-olympiad-programming-model/",
      "links": [
        "https://huggingface.co/NousResearch/NousCoder-14B",
        "https://nousresearch.com/nouscoder-14b-a-competitive-olympiad-programming-model/"
      ],
      "page_url": "https://nlp.com.ai/models/nouscoder-14b/",
      "record_updated_at": "2026-10-11T08:37:45.467886+00:00"
    },
    {
      "slug": "lfm-25-base-instruct-japanese-vision-language",
      "source_file": "2026-01-05-lfm-2.5.json",
      "release_date": "2026-01-05",
      "date": "January 5, 2026",
      "name": "LFM 2.5 (Base, Instruct, Japanese, Vision-Language)",
      "developer": "Liquid AI (USA)",
      "parameters": "1.2B / 1.5B / 1.6B (family variants with hybrid architecture)",
      "context_window": "",
      "license": "LFM Open License v1.0",
      "key_features": "Comprehensive on-device model family built on Liquid AI's hybrid LFM architecture and released as a unified LFM2.5 generation. Official release materials describe a scaled training recipe with extended pretraining from 10T to 28T tokens plus expanded post-training and reinforcement learning, targeting strong edge performance under tight memory budgets. The family includes text-generation checkpoints (LFM2.5-1.2B-Base, LFM2.5-1.2B-Instruct, LFM2.5-1.2B-JP), a vision-language checkpoint (LFM2.5-VL-1.6B), and an end-to-end audio-language checkpoint (LFM2.5-Audio-1.5B). Instruct benchmark tables report strong sub-2B results across GPQA, MMLU-Pro, IFEval, IFBench/Multi-IF, AIME25, and BFCLv3; JP release notes emphasize Japanese instruction-following gains; VL notes emphasize improved multi-image/OCR and multilingual visual understanding; Audio notes highlight speech/text I/O with a custom LFM-based detokenizer and lower-latency generation. Deployment support is published across Transformers, llama.cpp (GGUF), MLX, vLLM, and ONNX, with partner optimization notes for AMD and Nexa/Qualcomm NPU paths. Release framing focuses on private, fast, always-on edge intelligence across laptops, mobile devices, and embedded scenarios.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai",
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base",
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct",
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP",
        "https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B",
        "https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B",
        "https://arxiv.org/abs/2511.23404"
      ],
      "page_url": "https://nlp.com.ai/models/lfm-25-base-instruct-japanese-vision-language/",
      "record_updated_at": "2026-10-11T08:37:45.465582+00:00"
    },
    {
      "slug": "llama-33-8b-instruct-allura-community-reconstruction",
      "source_file": "2025-12-30-llama-3-3-8b-instruct.json",
      "release_date": "2025-12-30",
      "date": "December 30, 2025",
      "name": "Llama 3.3 8B Instruct (Allura community reconstruction)",
      "developer": "Allura (Forge) / Fizzarolli (reconstruction of a Meta model)",
      "parameters": "8,030,261,248 (8.03B; dense, including untied input embeddings and output head)",
      "context_window": "8,192 tokens configured in the original export; validated maximum unknown",
      "license": "Publisher-declared Llama 3.3 Community License Agreement + Acceptable Use Policy; custom restricted license, not Apache/MIT. Redistribution requires the agreement, Built with Llama attribution and NOTICE; distributed derived model names must begin with Llama. More than 700 million release-date monthly active users requires a separate Meta grant; no revenue threshold. Original repository lacks standalone LICENSE/NOTICE files; historical API terms and applicability to the reconstructed checkpoint remain review flags.",
      "key_features": "Community-published BF16, text-only autoregressive LlamaForCausalLM checkpoint, not a verified direct Meta public-weight release. Meta's April 2025 LlamaCon announcement confirms a new Llama 3.3 8B model for API fine-tuning, but that limited-preview service announcement is not this checkpoint's public-weight date. Allura's publisher reports downloading a merged fine-tuned model plus its adapter from Meta's API and subtracting the adapter to reconstruct the pre-adapter instruction model; the subtraction code, original export and numerical recovery accuracy were not independently verified. Export metadata reportedly names meta-llama/Meta-Llama-3-8B-Instruct, so behavioral differences and benchmark results do not by themselves prove exact Llama 3.3 ancestry or equivalence to the hosted model. The downloaded checkpoint has 32 dense Transformer layers, width 4096, SwiGLU intermediate width 14336, grouped-query attention with 32 query and 8 key/value heads, head dimension 128 and an untied 128256-token vocabulary. Original configuration sets 8192 positions, RoPE theta 500000 and no RoPE scaling; the tokenizer's huge model_max_length is a sentinel, and generation_config max_length 4096 is a generation default rather than the context ceiling. A Llama-header chat template is provided in a separate Jinja file. The publisher's OpenBench evaluation at temperature 1.0 reports IFEval 81.95 versus 78.2 for Llama 3.1 8B, averaging strict/loose instruction/prompt accuracies over one epoch, and GPQA Diamond 37.0 versus 29.3 over three epochs; these are community-reported measurements, not Meta's official benchmark results or independently reproduced scores. The publisher acknowledges factual hallucinations and withdrew tau-bench scores because looping traces and apparently incorrect successes made scoring unreliable. The exact checkpoint's pretraining corpus, token budget, instruction-tuning and reinforcement-learning recipe, distillation history and knowledge cutoff remain undisclosed. December 30 UTC is a provisional available-by date supported by a public download discussion, community conversion and local Q4 execution; those early events fall on December 29 in UTC-05, and exact first public access remains unresolved. This card covers only Allura's original 8K BF16 export. Bartowski's related GGUF packages and shb777's 128K RoPE/template/generation-config repackaging are separately documented; the latter has identical four shard hashes, not demonstrated new training or a verified native 128K Meta export.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/allura-forge/Llama-3.3-8B-Instruct",
        "https://about.fb.com/ltam/news/2025/04/todo-lo-que-anunciamos-en-nuestro-primer-llamacon/",
        "https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_3/LICENSE",
        "https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_3/USE_POLICY.md",
        "https://www.reddit.com/r/LocalLLaMA/comments/1pz7bmv/llama338binstruct/"
      ],
      "page_url": "https://nlp.com.ai/models/llama-33-8b-instruct-allura-community-reconstruction/",
      "record_updated_at": "2026-10-11T08:37:45.463131+00:00"
    },
    {
      "slug": "wedlm-8b",
      "source_file": "2025-12-29-wedlm-8b.json",
      "release_date": "2025-12-29",
      "date": "December 29, 2025",
      "name": "WeDLM-8B",
      "developer": "WeChat AI, Tencent (China)",
      "parameters": "8B dense; 8,190,735,360 learned parameters each (Base and Instruct; BF16)",
      "context_window": "32,768 tokens advertised; 16,384 configured",
      "license": "Custom Tencent WeDLM terms based on Apache 2.0, with an overriding EU territorial limitation; upstream Qwen Apache 2.0 and nano-vllm MIT notices",
      "key_features": "Text-only causal masked-diffusion Base and Instruct checkpoints: the developer identifies Qwen3-8B-Base as the initialization for WeDLM-8B-Base, followed by instruction post-training for WeDLM-8B-Instruct. The 36-layer dense transformer uses grouped-query attention (32 query heads, 8 KV heads), RoPE and untied input/output embeddings. Topological Reordering moves observed tokens to the physical prefix while retaining their logical RoPE position IDs, allowing masked-token recovery with strictly causal attention rather than bidirectional attention. Dual-Stream Masking trains a clean memory stream and block-wise masked prediction stream; the report describes 100B tokens of continued pretraining with a masked-denoising objective plus auxiliary autoregressive next-token loss, followed by supervised instruction fine-tuning. Its SFT quantities conflict between 10B tokens in the introduction and 10K internal instruction-response pairs in the training section; no checkpoint-specific RL or distillation stage is established. Streaming Parallel Decoding uses entropy-based acceptance, a positional distance penalty, immediate contiguous-prefix KV commitment and a dynamically refilled mask window; optimized generation requires the dedicated wedlm engine, while direct conversions retain ordinary AR generation without automatically gaining diffusion speedups. The report's ten-task Instruct average is 77.53 versus 75.12 for its Qwen3-8B-Instruct comparator, including GSM8K 92.27 (3-shot), HumanEval 80.49 (4-shot), MMLU 75.14 (5-shot) and GPQA-Diamond 44.95 (5-shot); MATH falls to 64.80 versus 69.60 and HellaSwag to 82.94 versus 86.13. These quality tables use step-wise one-token selection, not the aggressive parallel speed setting. The authors report a GSM8K parallel-decoding quality-speed frontier of 272–445 tokens/s, with accuracy declining from 92.3% to 79.7%; comparative GPU, batch size and engine version are not specified, so advertised task-dependent speed multipliers are not universal guarantees. The 32K model-card context conflicts with the saved 16K position configuration; tokenizer metadata says 128K and the engine defaults to 4K, neither establishing a validated checkpoint context extension. Released weights carry custom territorial terms despite Apache-2.0 labels in the cards.",
      "paper_url": "https://arxiv.org/abs/2512.22737",
      "links": [
        "https://huggingface.co/tencent/WeDLM-8B-Base",
        "https://huggingface.co/tencent/WeDLM-8B-Instruct",
        "https://github.com/Tencent/WeDLM",
        "https://wedlm.github.io"
      ],
      "page_url": "https://nlp.com.ai/models/wedlm-8b/",
      "record_updated_at": "2026-10-11T08:37:45.460655+00:00"
    },
    {
      "slug": "hyperclova-x-seed-8b-omni",
      "source_file": "2025-12-29-hyperclova-x-seed-8b-omni.json",
      "release_date": "2025-12-29",
      "date": "December 29, 2025",
      "name": "HyperCLOVA X SEED 8B Omni",
      "developer": "NAVER Cloud (South Korea)",
      "parameters": "8B nominal dense backbone; 8,590,233,600 text parameters; 13,272,225,599 stored safetensor elements across backbone, encoders, projectors and vision decoders, excluding audio-decoder archives (not a verified unique learned total)",
      "context_window": "32K advertised shared context; saved text-decoder configuration: 8,192 positions (discrepancy unresolved)",
      "license": "HyperCLOVA X SEED 8B Omni Model License Agreement + Acceptable Use Policy (custom, not Apache 2.0). Separate NAVER permission required for release-time preceding-month MAU >10 million or substantially similar/competing products; redistribution, Powered by HyperCLOVA X attribution and derivative-model naming obligations apply. Upstream component provenance/terms require review.",
      "key_features": "Korean-first any-to-any pathfinding model with a single dense autoregressive backbone jointly predicting interleaved text, discrete image and discrete speech tokens, supplemented by continuous modality encoders and signal decoders. Accepts text, images, video and audio; generates text, images and speech, not video. The 36-layer, 4,096-wide decoder uses grouped-query attention with 32 query/8 KV heads and an untied 200,704-token vocabulary. Visual understanding uses a Qwen2.5-VL-style encoder and linear adapter; generation uses a frozen SigLIP2-derived TA-Tok tokenizer and a separately trained diffusion decoder in FLUX.1 VAE latent space, with a smaller autoguidance model. Audio uses a Qwen2-Audio/Whisper-large-v3-derived continuous encoder, a SenseVoice/CosyVoice-derived discrete tokenizer and Unit-BigVGAN waveform decoders; MambaMia compresses video audio from 25 Hz to 1 Hz. Its own report describes text pretraining at 4K/8K/32K with auxiliary multi-token prediction, then 302B-token modality-vocabulary alignment, 2.3T-token joint discrete-modality pretraining, approximately 20B-token long-context adaptation, and continuous-vision alignment plus 1.5T-token multimodal training and audio-adapter alignment. Four supervised post-training stages cover foundational dialogue, task specialization, video/long context and intent-aware multistep reasoning; no released-policy RL recipe is established. Exact initial text checkpoint, optimizer and total text-pretraining budget are undisclosed. The January 5 technical report evaluates TAG-2025-12-31, whose safetensor hashes match the preserved December 29 snapshot: KMMLU-Pro 64.9, MMLU-Pro 54.2, GSM8K 87.3, GenEval 0.64 and ImgEdit 3.83; Video-MME scores 58.2 without subtitles, below Qwen2.5-Omni-7B's 64.3. Korean TTS MOS is 4.22 ± 0.11 in a 30-listener, 20-utterance internal study. These are multimodal-system results, not independently reproduced bare-backbone scores. OmniServe exposes an OpenAI-compatible interface and requires an image-generation tool definition and output storage; native modules should not be confused with the 32B Think demo's external speech pipeline. The local template opens a think block unless skip_reasoning=true; service defaults and output budgets are separate from context capacity. Original core weights are F32, with BF16 vision-decoder files; the advertised 8B/approximately 16GB description is not the full downloadable package footprint.",
      "paper_url": "https://arxiv.org/abs/2601.01792",
      "links": [
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Omni-8B",
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Omni-8B/tree/TAG-2025-12-31",
        "https://github.com/NAVER-Cloud-HyperCLOVA-X/OmniServe",
        "https://www.navercorp.com/en/media/pressReleasesDetail?seq=34256"
      ],
      "page_url": "https://nlp.com.ai/models/hyperclova-x-seed-8b-omni/",
      "record_updated_at": "2026-10-11T08:37:45.458668+00:00"
    },
    {
      "slug": "hyperclova-x-seed-32b-think",
      "source_file": "2025-12-29-hyperclova-x-seed-32b-think.json",
      "release_date": "2025-12-29",
      "date": "December 29, 2025",
      "name": "HyperCLOVA X SEED 32B Think",
      "developer": "NAVER Cloud (South Korea)",
      "parameters": "32B dense; 33,313,410,304 total checkpoint parameters (32,598,348,800 language model + 688,841,984 vision encoder + 26,219,520 projector)",
      "context_window": "128K tokens (131,072 configured; shared text/visual context)",
      "license": "HyperCLOVA X SEED 32B Think Model License Agreement + Acceptable Use Policy (custom; conditional commercial use, attribution/naming obligations; separate permission required for >10M monthly active users at release or NAVER-competing products/services)",
      "key_features": "Korean-centric vision-language reasoning and tool-using agent model with text, image, and video inputs and text output; speech interaction in NAVER’s omnimodal demonstrations uses external speech recognition/synthesis modules, not native audio weights in this checkpoint. The 72-layer, 5,120-wide decoder uses grouped-query attention with 40 query heads and 8 KV heads, RMSNorm, SwiGLU, and untied input/output embeddings. A 32-layer Qwen2.5-VL-style vision Transformer with 2×14×14 spatiotemporal patch embedding and window attention feeds a linear projector into the text embedding space; this is not the earlier SEED 3B SigLIP/C-Abstractor architecture. The released image processor sets a 2,073,600-pixel maximum; the report describes efficient image/video budgets of roughly 3K/11K visual tokens using up to 120 video frames, but its video sampling settings are not fully encoded in the original exported video processor. Own four-stage Korean/English pretraining uses quality filtering, synthetic reasoning data, 10% fill-in-the-middle tokens, and context progression from 4K through 8K and 32K; the initial 4K stage alone consumes 6T tokens, not the entire training run. Text SFT is followed by four-stage multimodal SFT with text replay, then multimodal verifiable-reward and agent RL using GRPO variants, and preference alignment using a learned reward model and PPO. The PPO stage trains on final responses rather than intermediate thinking paths; neither direct initialization from the released 14B Think weights nor a particular external base checkpoint is established. Supports reasoning/non-reasoning, multi-turn dialogue, and XML-format tool calls; the original local template opens <think> by default and uses skip_reasoning=true to close it, distinct from OmniServe’s documented thinking=false and 500-token service budget. In the dedicated report, CLIcK scores 75.2, HAERAE-1.0 87.4, and TextVQA 85.4; Tau² Airline/Retail/Telecom scores are 58.0/71.6/65.1 with a GPT-4.1 user simulator, reasoning enabled, and up to 200 turns, while TerminalBench 1.0 scores 21.8 using Terminus-2. These are reported evaluation-system results, not independently reproduced scores for the exported tensors. NAVER reports Grade 1 results across major 2026 KCSAT subjects from image inputs, including perfect English and Korean-history scores. The report notes weaker English-centric performance against some baselines and text-performance degradation during sequential multimodal training; original processor/template integration also required compatibility fixes. Separate from earlier 14B Think, April SEED models, and the co-announced native 8B Omni.",
      "paper_url": "https://arxiv.org/abs/2601.03286",
      "links": [
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-32B",
        "https://github.com/NAVER-Cloud-HyperCLOVA-X/OmniServe",
        "https://www.navercorp.com/en/media/pressReleasesDetail?seq=34256"
      ],
      "page_url": "https://nlp.com.ai/models/hyperclova-x-seed-32b-think/",
      "record_updated_at": "2026-10-11T08:37:45.456441+00:00"
    },
    {
      "slug": "minimax-m21",
      "source_file": "2025-12-26-minimax-m2-1.json",
      "release_date": "2025-12-26",
      "date": "December 26, 2025",
      "name": "MiniMax-M2.1",
      "developer": "MiniMax (China)",
      "parameters": "230B total / ~10B active (MoE; 228,689,748,992 learned parameters in the released checkpoint)",
      "context_window": "196,608 tokens",
      "license": "Modified MIT: retain copyright/license notices and prominently display “MiniMax M2.1” in the UI of any commercial product or service using the software or derivatives; no revenue or MAU threshold. Launch-linked December 22, 2025 license is unchanged in the checked GitHub history; no standalone HF LICENSE file.",
      "key_features": "Text-only reasoning and agentic coding successor to MiniMax-M2, released as a mixed FP8/BF16/F32 checkpoint. The 62-layer decoder uses full softmax attention with 48 query heads and 8 KV heads, a 3,072-wide hidden state, and 256 routed experts with 8 selected per token and no shared experts; quantized weights use 128×128 FP8 blocks. Although its config enables multi-token prediction, the released tensor inventory contains no MTP heads. Focuses on multilingual software engineering, native Android/iOS and full-stack web development, composite instruction following, office workflows, and more concise reasoning than M2. The publisher’s January 2026 account identifies further post-training over M2: multi-scaffold rejection-sampled SFT and agentic CISPO RL in Forge, executable GitHub PR tasks, expert-designed AppDev rubrics verified through interactive agents, and WebExplorer-style long-horizon search synthesis; it reports 10,000+ runnable PRs and 140,000+ variable tasks as of M2.1, not a disclosed complete launch training budget. The chat template starts generation with <think>, retains reasoning during the current user/tool turn, and strips earlier-turn reasoning; XML tool calls require external execution, and no built-in hard thinking budget or non-thinking switch is documented. Publisher-reported SWE-bench Verified 74.0, Multi-SWE-bench 49.4 and SWE-bench Multilingual 72.5 use Claude Code with an overridden system prompt and four-run means; Droid and mini-swe-agent SWE scores are 71.3 and 67.0. Terminal-Bench 2.0 47.9 uses an internal framework with timeout limits removed; VIBE 88.6 is a three-run agent-verified application-development score. These are scaffold-dependent release claims, not independently reproduced scores or proof of exact hosted/downloaded checkpoint equivalence. Improvements are not universal: reported LCB 81.0 and IFBench 70.0 are below M2’s 83.0 and 72.0. Supported local sequence length is distinct from the tokenizer’s 40,960,000 ceiling and the hosted API’s 204,800-token limit. API/Agent launch preceded public-weight release; December 26 is verified availability-by, while the exact first public unlock remains unresolved. Lightning/highspeed is a hosted throughput mode, not a separately verified downloadable checkpoint.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M2.1",
        "https://modelscope.cn/models/MiniMax/MiniMax-M2.1",
        "https://github.com/MiniMax-AI/MiniMax-M2.1",
        "https://www.minimax.io/news/minimax-m21",
        "https://www.minimax.io/news/post-training-experience-and-insights-for-agent-models",
        "https://github.com/MiniMax-AI/MiniMax-M2.1/blob/ea93fc40128a5bc366e3531a924f01f3d6d9cc15/LICENSE",
        "https://x.com/MiniMax_AI/status/2004524661359407129"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m21/",
      "record_updated_at": "2026-10-11T08:37:45.454414+00:00"
    },
    {
      "slug": "lfm2-26b-exp",
      "source_file": "2025-12-25-lfm2-2-6b-exp.json",
      "release_date": "2025-12-25",
      "date": "December 25, 2025",
      "name": "LFM2-2.6B-Exp",
      "developer": "Liquid AI (USA)",
      "parameters": "2.6B (2,569,272,320 unique learned parameters; dense, tied embeddings)",
      "context_window": "32,768 tokens",
      "license": "LFM Open License v1.0 (custom revenue-limited commercial license; not Apache 2.0)",
      "key_features": "Experimental text-only LFM2 checkpoint, continuing LFM2-2.6B with RL-only post-training focused on instruction following, knowledge, and math; the exact RL algorithm, reward construction, training budget, and initialization stage within the parent pipeline are not disclosed. Its 30-layer hybrid backbone combines 22 double-gated, kernel-size-3 depthwise causal convolution blocks with 8 full grouped-query attention blocks (32 query heads, 8 KV heads), width 2,048, SwiGLU feed-forward layers, RMSNorm, and RoPE. Input-dependent multiplicative gates are applied before and after each short convolution; attention supplies long-range mixing while convolution uses a fixed-size local cache. Advertised context is 32,768 tokens, distinct from the saved 128,000-position configuration and unbounded tokenizer sentinel. The publisher reports IFBench 44.40, Multi-IF 73.05, IFEval 88.13, MMLU-Pro 53.87, GPQA 41.31, and AIME25 22.67; these chart scores have no disclosed sampling or scaffold protocol, and its claimed IFBench advantage over DeepSeek R1-0528 is benchmark-specific. The original release stored F32 tensors; the publisher re-exported BF16 weights on December 26, and provides F16 and quantized GGUF packages for llama.cpp. Uses a ChatML-like format and supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. Designed for phones and laptops, with recommended narrow-use-case fine-tuning for agentic workflows, extraction, RAG, creative writing, and multi-turn conversations; the publisher nevertheless discourages programming and knowledge-intensive applications. Later documentation marks this experimental checkpoint deprecated in favor of LFM2.5-2.6B; the successor is a separate release.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2-2.6B-Exp",
        "https://huggingface.co/LiquidAI/LFM2-2.6B-Exp-GGUF",
        "https://x.com/liquidai/status/2004190178068296181",
        "https://docs.liquid.ai/lfm/models/lfm2-2.6b-exp"
      ],
      "page_url": "https://nlp.com.ai/models/lfm2-26b-exp/",
      "record_updated_at": "2026-10-11T08:37:45.452238+00:00"
    },
    {
      "slug": "glm-47",
      "source_file": "2025-12-22-glm-4-7.json",
      "release_date": "2025-12-22",
      "date": "December 22, 2025",
      "name": "GLM-4.7",
      "developer": "Z.ai / Zhipu AI (China)",
      "parameters": "MoE: nominal 355B total / 32B active; 352.798B main-model learned parameters + 5.540B exported MTP parameters (358.338B packaged, including shared-weight copies)",
      "context_window": "202,752 tokens (released checkpoint configuration; hosted service advertises 200K)",
      "license": "MIT (publisher declarations for BF16 and FP8 weights; standalone weight LICENSE absent); official repository code: Apache 2.0",
      "key_features": "Text-only hybrid thinking/non-thinking flagship with original BF16 and official mixed-FP8 packages; FP8 same-launch grouping remains provisional pending independent public-access confirmation. The released decoder has 92 layers, hidden size 5,120, full-attention GQA with 96 query heads and 8 KV heads, QK normalization, three initial dense layers, and 160 routed experts selecting 8 plus one shared expert per MoE layer; it includes a separately exported one-layer multi-token prediction module. FP8 uses compressed-tensors per-channel weight scales and dynamic per-token activation quantization, with selected tensors retained at higher precision. Interleaved thinking precedes responses and tool calls; turn-level control uses enable_thinking=false for direct responses. The released template enables thinking by default but clears reasoning from earlier user turns unless clear_thinking=false is explicitly supplied; preserved thinking requires retained reasoning_content and was recommended for agentic tasks with SGLang at launch. No checkpoint-level hard thinking-token budget is encoded in the template. Both original precision configs set max_position_embeddings=202752 without a RoPE-scaling override; tokenizer model_max_length=128000 is a separate saved default, and the hosted 128K maximum output is not an additional context window. Publisher-reported original GLM-4.7 results include SWE-bench Verified 73.8%, SWE-bench Multilingual 66.7%, Terminal Bench 2.0 41.0%, LiveCodeBench v6 84.9%, AIME 2025 95.7%, HMMT February/November 2025 97.1%/93.5%, HLE 24.8% without tools and 42.8% with tools, tau-squared-Bench 87.4%, and BrowseComp 52.0% or 67.5% with context management. These are model-plus-evaluation-system claims, not independently verified BF16/FP8 equivalence or unscaffolded pass@1: launch settings use up to 131072 new tokens for most tasks, 16384 for Terminal/SWE, preserved thinking for multi-turn agents, and tau-squared user-prompt/domain fixes; attempt aggregation and complete agent scaffolds are undisclosed. The publisher describes stronger multilingual agentic coding, cleaner web UI and slide layouts, more vivid creative writing and role-play, and better multi-turn consistency, with integration into Claude Code, Kilo Code, Cline and Roo Code. GLM-4.7-specific pretraining/continued-training token budgets, data mixtures and complete SFT/RL recipe were not disclosed in the checked release sources; the cited GLM-4.5 report does not document this checkpoint. Later GLM-4.7-Flash, hosted FlashX and third-party quantizations are not part of this December grouping.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/zai-org/GLM-4.7",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.7",
        "https://huggingface.co/zai-org/GLM-4.7-FP8",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.7-FP8",
        "https://z.ai/blog/glm-4.7",
        "https://github.com/zai-org/GLM-4.5",
        "https://docs.z.ai/guides/llm/glm-4.7",
        "https://arxiv.org/abs/2508.06471"
      ],
      "page_url": "https://nlp.com.ai/models/glm-47/",
      "record_updated_at": "2026-10-11T08:37:45.450477+00:00"
    },
    {
      "slug": "t5gemma-2",
      "source_file": "2025-12-18-t5gemma-2.json",
      "release_date": "2025-12-18",
      "date": "December 18, 2025",
      "name": "T5Gemma 2",
      "developer": "Google (USA)",
      "parameters": "Three dense pretrained sizes: 270M-270M (786,029,296 total), 1B-1B (2,115,977,456 total), and 4B-4B (7,508,928,880 total), including the frozen vision encoder and connector; shared embeddings counted once",
      "context_window": "128K input and up to 32K output advertised for all sizes; configured encoder/decoder positional ceilings separately: 32,768/32,768 for 270M-270M and 1B-1B, 131,072/131,072 for 4B-4B",
      "license": "Gemma Terms of Use with the Gemma Prohibited Use Policy (custom weights license, not Apache 2.0)",
      "key_features": "Gemma 3-adapted multimodal encoder-decoder research family: three UL2-pretrained checkpoints, not released instruction-tuned models or separate PrefixLM variants. Both text branches reuse the corresponding Gemma 3 pretrained weights rather than training from scratch; all sizes accept text and images and generate text, including the two sizes whose Gemma 3 parents were text-only. A frozen 416.9M-parameter SigLIP vision encoder processes 896x896 images into 256 tokens, with a learned connector feeding the text encoder. Encoder input, decoder input and output word embeddings are tied; decoder self- and cross-attention share a merged attention module, avoiding a separate cross-attention parameter stack. Dense gated-GELU Transformers use GQA, QK normalization, RMSNorm, RoPE and five local-attention layers per global layer; attention projections have no bias. Continued pretraining uses approximately 2T multilingual web, code, mathematics and image tokens, five text-denoising tasks plus image-prefix language modeling, cross-entropy, and separate input/output sequences up to 16K; pretraining distillation was explicitly dropped, and the final model averages the last five checkpoints. Supports over 140 languages. Publisher reports improved multimodal and long-context capabilities: pretrained 4B-4B scores 39.4 on MMMU and 81.7/57.6 on RULER at 32K/128K, but not every task improves over Gemma 3. The 128K small-model input claim exceeds their saved 32,768 positional settings; the required inference override is not documented in the checked release sources. Paper post-training results are illustrative lightweight distillation/SFT experiments without RL, not scores for released IT checkpoints. Pretrained weights are intended for downstream adaptation before deployment; factual errors remain possible, and safety evaluation used English prompts only.",
      "paper_url": "https://arxiv.org/abs/2512.14856v1",
      "links": [
        "https://huggingface.co/collections/google/t5gemma-2",
        "https://huggingface.co/google/t5gemma-2-270m-270m",
        "https://huggingface.co/google/t5gemma-2-1b-1b",
        "https://huggingface.co/google/t5gemma-2-4b-4b",
        "https://www.kaggle.com/models/google/t5gemma-2",
        "https://blog.google/technology/developers/t5gemma-2",
        "https://ai.google.dev/gemma/terms"
      ],
      "page_url": "https://nlp.com.ai/models/t5gemma-2/",
      "record_updated_at": "2026-10-11T08:37:45.448242+00:00"
    },
    {
      "slug": "hearthfire-24b",
      "source_file": "2025-12-18-hearthfire-24b.json",
      "release_date": "2025-12-18",
      "date": "December 18, 2025",
      "name": "Hearthfire-24B",
      "developer": "Latitude (USA)",
      "parameters": "24B (dense; 23,572,403,200 parameters)",
      "context_window": "131,072 tokens (configured; finetune maximum not independently validated)",
      "license": "Apache 2.0 (publisher declaration; standalone LICENSE file absent)",
      "key_features": "Text-only narrative-writing SFT finetune of Mistral Small 3.2-24B-Instruct-2506, developed with Gryphe Padar; the released checkpoint contains the text decoder, not the parent's vision encoder or multimodal connector. Uses 40 decoder layers, grouped-query attention with 32 query and 8 key/value heads, and untied input/output embeddings; released as BF16 safetensors, with publisher-provided GGUF downloads also available. Trained on a single dataset of several thousand longform 8–16K writing examples, predominantly second-person present tense with approximately 10% third-person examples; continue-heavy narratives interleave extended prose with sparse player actions. Prioritizes atmosphere, introspection and slow-burn pacing over constant external conflict; the publisher describes a warm, cooperative tone and explicitly permits the model to act and speak for the player to maintain narrative flow. ChatML prompt format; recommended temperature 0.8, repetition penalty 1.05 and min_p 0.025. The 131,072-token positional configuration matches the direct parent rather than establishing a new context extension, and the training-example lengths are not a validated inference limit. Launch generation configuration stopped on token 2 rather than ChatML end token 999; the publisher merged a December 19 fix to stop on both. No checkpoint-specific benchmark results or DPO/RL stages are reported. Available to try free on AI Dungeon.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Hearthfire-24B",
        "https://huggingface.co/LatitudeGames/Hearthfire-24B-GGUF",
        "https://x.com/kolbytn/status/2001781551777092077"
      ],
      "page_url": "https://nlp.com.ai/models/hearthfire-24b/",
      "record_updated_at": "2026-10-11T08:37:45.446502+00:00"
    },
    {
      "slug": "functiongemma-270m",
      "source_file": "2025-12-18-functiongemma.json",
      "release_date": "2025-12-18",
      "date": "December 18, 2025",
      "name": "FunctionGemma 270M",
      "developer": "Google (USA)",
      "parameters": "270M nominal; 268,098,176 total dense parameters, including 167,772,160 tied embedding parameters",
      "context_window": "32K (32,768 tokens shared between input and generated output)",
      "license": "Gemma Terms of Use + Gemma Prohibited Use Policy (custom open-weight terms; not Apache 2.0)",
      "key_features": "Text-only, decoder-only Gemma 3 270M derivative fine-tuned specifically for function calling, released as google/functiongemma-270m-it in BF16 with Google-endorsed llama.cpp GGUF packaging. Uses 18 layers, width 640, grouped-query attention with four query heads and one KV head, and a 262,144-token vocabulary; fifteen 512-token sliding-attention layers alternate in five-to-one groups with three full-attention layers. Designed as a lightweight foundation for customized local agents, not a general-purpose direct dialogue model. Trained on public API definitions and tool-use interactions containing prompts, function calls, function responses and natural-language responses; the exact FunctionGemma adaptation objective, token budget and PT-versus-IT initialization checkpoint are not disclosed. Explicit training scope covers single-turn and parallel calls, not out-of-the-box multi-turn state tracking or dependent multi-step chaining; task-specific fine-tuning is recommended. Its custom developer/user/model chat format serializes tool declarations, calls and results using function-control tokens and <escape>-delimited strings rather than ordinary JSON output. The application must parse generated calls, execute tools and feed results back; neither the template nor tool training guarantees constrained decoding or automatic runtime execution. Google reports zero-shot BFCL Simple 61.6, Multiple 63.5, Parallel 39.0 and Parallel Multiple 29.5, with lower Live-subset results; these are individual subsets, not an overall BFCL score. On a held-out Mobile Actions evaluation, the released foundation scores 58%, rising to 85% after additional task-specific fine-tuning, not 85% for the unadapted checkpoint. A public completion-only supervised fine-tuning recipe supports customization; the separately adapted Tiny Garden LiteRT-LM demo illustrates offline game control. Google's mobile latency and memory figures concern dynamic-int8 specialized demos on a Samsung S25 Ultra, not the BF16 foundation. Supports local-first deployments and routing to larger models through an application scaffold. December 18 is verified availability-by through Google's explicit download offer and contemporaneous local-installation testimony; exact first public unlock remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/google/functiongemma-270m-it",
        "https://huggingface.co/ggml-org/functiongemma-270m-it-GGUF",
        "https://blog.google/innovation-and-ai/technology/developers-tools/functiongemma/",
        "https://ai.google.dev/gemma/docs/functiongemma/model_card",
        "https://ai.google.dev/gemma/docs/functiongemma/formatting-and-best-practices",
        "https://www.kaggle.com/models/google/functiongemma",
        "https://ai.google.dev/gemma/terms",
        "https://ai.google.dev/gemma/prohibited_use_policy"
      ],
      "page_url": "https://nlp.com.ai/models/functiongemma-270m/",
      "record_updated_at": "2026-10-11T08:37:45.443935+00:00"
    },
    {
      "slug": "nemotron-cascade",
      "source_file": "2025-12-16-nemotron-cascade.json",
      "release_date": "2025-12-16",
      "date": "December 16, 2025",
      "name": "Nemotron-Cascade",
      "developer": "NVIDIA (USA)",
      "parameters": "8B / 14B dense (8,190,735,360 parameters each for 8B unified and 8B-Thinking; 14,768,307,200 for 14B-Thinking, including untied input/output embeddings)",
      "context_window": "32,768 tokens native/configured; explicit YaRN extension to 64K for all three, and 90K for 14B-Thinking SWE evaluation",
      "license": "NVIDIA Open Model License Agreement (custom; linked October 24, 2025 version) + NVIDIA Trustworthy AI terms; preserve applicable Apache-2.0 Qwen3 ancestry notices/obligations",
      "key_features": "Original three-checkpoint, text-only Qwen3-8B-Base/Qwen3-14B-Base post-training family, not the later Nemotron-Cascade 2 MoE or Nemotron 3 hybrid architecture. BF16 dense causal Transformers use 36/40 layers and grouped-query attention with 32/40 query heads and 8 KV heads. The unified 8B supports per-user-turn /think and /no_think control; the two Thinking variants are thinking-only, with ChatML and tool-call templates. Two one-epoch SFT stages at 16K then 32K combine general conversation, math, code and science with later tool-use and software-engineering data; DeepSeek-R1 and R1-0528 supply reasoning traces, with V3-0324 non-thinking and Qwen3-235B tool-call data. Sequential Cascade RL then applies RLHF, instruction-following RL, math RL, code RL and SWE RL using strictly on-policy GRPO without KL regularization. A Qwen2.5-72B-Instruct-derived preference reward model drives RLHF; dedicated Thinking instruction-following combines rule and preference rewards, math uses answer verification, code uses test execution, and SWE uses execution-free lexical/semantic patch similarity with Kimi-Dev-72B. NVIDIA reports reduced engineering complexity and limited cross-domain forgetting; training cost is not disclosed, while one code-verification setup reduces batch verification latency from 1172.4 to 416.2 seconds through asynchronous computation. Final 14B-Thinking reports LiveCodeBench v5/v6 scores of 77.5/74.6, exceeding its 671B R1-0528 SFT teacher; unified 8B reports 74.3/71.1. Reported pass@1 scores average multiple samples, rather than selecting the best sample. SWE-bench Verified is 43.1 for 14B and 37.2 for unified 8B without test-time scaling, rising to 53.8/43.6 with Agentless patch validation and selection. IOI 2025 silver-medal-level 343.37 is a feedback-driven evaluation pipeline, not a standalone-model medal: 128K thinking budget, 20 candidates per subtask per round, sample-test filtering, Tail-10 selection and at most 50 official-judge submissions per problem. The 64K normal evaluation output budget and IOI 128K thinking budget are not native context declarations. Strengths are coding, mathematics and transparent released recipes/data; 14B scores remain below R1-0528 on knowledge reasoning, AIME and SWE, and BFCL V3 67.5 trails Qwen3-14B 70.4. Original public-release day remains provisional: publisher page says December 15, checkpoint READMEs say December 8, and all three have independent weight-conversion evidence by December 16 UTC; upload dates do not establish first public access. Later December 19 SFT/RL intermediate checkpoints are separate release-date candidates, not silently grouped here.",
      "paper_url": "https://arxiv.org/abs/2512.13607v1",
      "links": [
        "https://huggingface.co/nvidia/Nemotron-Cascade-8B",
        "https://huggingface.co/nvidia/Nemotron-Cascade-8B-Thinking",
        "https://huggingface.co/nvidia/Nemotron-Cascade-14B-Thinking",
        "https://modelscope.cn/models/nv-community/Nemotron-Cascade-8B",
        "https://modelscope.cn/models/nv-community/Nemotron-Cascade-8B-Thinking",
        "https://modelscope.cn/models/nv-community/Nemotron-Cascade-14B-Thinking",
        "https://huggingface.co/collections/nvidia/nemotron-cascade",
        "https://research.nvidia.com/labs/nemotron/nemotron-cascade",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-cascade/",
      "record_updated_at": "2026-10-11T08:37:45.441454+00:00"
    },
    {
      "slug": "mimo-v2-flash-base-and-post-trained",
      "source_file": "2025-12-16-mimo-v2-flash.json",
      "release_date": "2025-12-16",
      "date": "December 16, 2025",
      "name": "MiMo-V2-Flash (Base and post-trained)",
      "developer": "Xiaomi (China)",
      "parameters": "309B total / 15B active (MoE); 308,778,780,864 main-model parameters + 987,820,224 across three MTP blocks",
      "context_window": "256K (262,144 tokens)",
      "license": "MIT (model weights; GitHub repository code LICENSE is Apache-2.0)",
      "key_features": "Original text-only Base and hybrid-thinking assistant checkpoints, released in mixed FP8 precision with three lightweight dense MTP blocks. The 48-layer, width-4096 Transformer uses 39 sliding-window GQA layers (128-token window; 64 query/8 KV heads) and nine global GQA layers (64 query/4 KV heads), rather than MLA or DeepSeek Sparse Attention; the first layer has a dense FFN, while the other 47 use 256 routed experts with eight selected per token and no shared experts. Learnable attention sinks stabilize SWA; the advertised 5:1 pattern has a first-layer global-attention exception. Randomly initialized pretraining uses AdamW and 27T tokens: 22T general pretraining, 4T code-upsampled mid-training with about 5% synthetic reasoning data, and 1T context extension from 32,768 to 262,144 tokens. One MTP head is used in pretraining, then replicated and jointly trained for multi-step prediction during post-training. Post-training combines millions of thinking/non-thinking SFT examples generated by in-house specialized checkpoints, domain-specialized RL/SFT teachers, and Multi-Teacher On-Policy Distillation (MOPD), using student rollouts with token-level reverse-KL guidance and outcome rewards. Code-agent RL scales to about 120K environments; video-based web-development rewards come from an external multimodal verifier, not native visual input. Launch-reported results include AIME 2025 94.1%, GPQA-Diamond 83.7%, SWE-bench Verified 73.4%, SWE-bench Multilingual 71.7%, and LongBench V2 60.6%. Coding and search scores require external tools/agent loops; BrowseComp rises from 45.4 to 58.3 with external context management, not a larger native context. Attempt budgets and score aggregation are incompletely disclosed. The report notes weaker Base factual knowledge and a remaining gap to the strongest closed models. Three-layer MTP enables self-speculative decoding; the launch blog reports 2.0–2.6× speedups, while the joint SGLang benchmark reports about 150 tokens/s per request on H200 at batch 16 per DP rank and up to 64K input, not a hardware-independent speed. Launch API pricing of $0.10/$0.30 per million input/output tokens is separate from local inference cost. Tool templates support thinking toggles and require preserving reasoning history in multi-turn thinking/tool use.",
      "paper_url": "https://github.com/XiaomiMiMo/MiMo-V2-Flash/blob/65f0e73804b030611b74235ddad9dc4ad1a6a586/paper.pdf",
      "links": [
        "https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash",
        "https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash-Base",
        "https://modelscope.cn/models/XiaomiMiMo/MiMo-V2-Flash",
        "https://modelscope.cn/models/XiaomiMiMo/MiMo-V2-Flash-Base",
        "https://github.com/XiaomiMiMo/MiMo-V2-Flash",
        "https://mimo.xiaomi.com/mimo-v2-flash",
        "https://lmsys.org/blog/2025-12-16-mimo-v2-flash/"
      ],
      "page_url": "https://nlp.com.ai/models/mimo-v2-flash-base-and-post-trained/",
      "record_updated_at": "2026-10-11T08:37:45.439217+00:00"
    },
    {
      "slug": "qwenlong-l15-30b-a3b",
      "source_file": "2025-12-15-qwenlong-l1-5.json",
      "release_date": "2025-12-15",
      "date": "December 15, 2025",
      "name": "QwenLong-L1.5-30B-A3B",
      "developer": "Tongyi Lab, Alibaba Group (China)",
      "parameters": "30.532B total; approximately 3.353B active per token (MoE, including input/output embeddings)",
      "context_window": "262,144 tokens (256K native); 1M–4M document processing requires iterative memory-agent inference",
      "license": "Apache 2.0 (publisher-declared; standalone checkpoint LICENSE files absent)",
      "key_features": "Text-only reasoning checkpoint post-trained from Qwen3-30B-A3B-Thinking-2507, with 48 Transformer layers, grouped-query attention (32 query/4 KV heads), and 128 routed experts selecting eight per token; released in BF16. Long-context data synthesis builds multi-hop questions from atomic document facts and globally distributed evidence, including corpus-level numerical reasoning and dialogue memory; 14.1K training examples survive filtering from 42.7K synthesized examples. Purely on-policy, token-normalized GRPO uses task-balanced sampling, task-specific reward-standard-deviation normalization, and Adaptive Entropy-Controlled Policy Optimization (AEPO), which switches negative-advantage masking on and off at upper/lower entropy thresholds. Rewards combine ground-truth containment checks with gpt-oss-120b judging when rule checks fail. Three full-context RL stages increase input/output budgets from 20K/12K to 60K/20K and 120K/50K; a separately trained memory-management expert is SCE-merged with Stage 3, followed by a fourth full-context RL stage. This produces one released checkpoint, not separate full-context and memory models: the memory agent repeatedly updates compact memory and a navigational plan across document chunks rather than expanding native attention to millions of tokens. The report averages 71.82 across six long-context benchmarks versus 61.92 for its parent, with MRCR 82.99 and CorpusQA 81.25; these evaluations cap input at 128K and generation at 50K, use middle truncation, and combine answer matching with model judging for QA. The abstract claims a 9.48-point ultra-long-task gain over the parent memory agent, with aggregation unspecified; Table 9 reports only 14.29 on CorpusQA at 4M tokens. LongMemEval rises from 60.80 to 76.40 and AIME25 from 82.81 to 86.46, while AIME24 and BFCL-V4 Memory-Rec_Sum slightly regress. Multilingual capability is inherited, but predominantly English post-training may weaken non-English long-context reasoning. Authors identify long-form-output coverage and subjective-task reward quality as limitations; the training dataset was not publicly released.",
      "paper_url": "https://arxiv.org/abs/2512.12967v1",
      "links": [
        "https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1.5-30B-A3B",
        "https://modelscope.cn/models/iic/QwenLong-L1.5-30B-A3B",
        "https://github.com/Tongyi-Zhiwen/Qwen-Doc/tree/main/QwenLong-L1.5"
      ],
      "page_url": "https://nlp.com.ai/models/qwenlong-l15-30b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.437066+00:00"
    },
    {
      "slug": "nemotron-3-nano-30b-a3b",
      "source_file": "2025-12-15-nemotron-3-nano.json",
      "release_date": "2025-12-15",
      "date": "December 15, 2025",
      "name": "Nemotron 3 Nano 30B-A3B",
      "developer": "NVIDIA (USA)",
      "parameters": "31.58B total; approximately 3.6B active including embeddings (3.2B under NVIDIA’s embedding-excluded convention); Base and post-trained checkpoints, with BF16 and post-trained FP8 packaging",
      "context_window": "Post-trained: up to 1M tokens in report/current support, 262,144 configured default, but pinned launch BF16 card declares 128K input; Base: 128K declared input/output versus 262,144 configuration and launch retrieval tests through 512K; family-level 1M scope unresolved for Base",
      "license": "Custom NVIDIA Open Model License Agreement at launch (October 24, 2025 version, including Trustworthy AI and guardrail conditions); subsequently NVIDIA Nemotron Open Model License (December 15, 2025 version); commercial use permitted, not Apache-2.0",
      "key_features": "Text-only, from-scratch hybrid Mamba-2–Transformer MoE model for reasoning, coding, tool use and agentic workflows. Its 52 sequential residual blocks comprise 23 Mamba-2, 23 MoE and 6 grouped-query attention blocks (32 query heads, 2 KV heads), not parallel Mamba/attention branches. Each MoE block routes to 6 of 128 experts and adds one shared MLP of twice the routed-expert width; no exported multi-token-prediction module. Base pretraining uses 25T processed tokens in a two-phase quality curriculum, followed by 121B tokens of long-context continued pretraining mixing 512K and 4K sequences; the disclosed corpus contains about 13.34T tokens, with more than 3T new unique tokens over Nemotron 2. The post-trained checkpoint adds supervised fine-tuning over 18M samples with 256K packing, synchronous multi-environment GRPO for verifiable math, code, science, instruction following and tool-use rewards, then GenRM-guided RLHF with group-relative length control. Synthetic teacher data includes Qwen, GPT-OSS and DeepSeek-generated traces; the weights are not a compressed Qwen checkpoint. DPO tool-hallucination experiments in the report were not used in the released model. Reasoning on/off is controlled through chat-template prefixes; finite reasoning budgets require generation-side stopping/continuation logic rather than a hard checkpoint switch. Supports English, German, Spanish, French, Italian and Japanese in post-training. Reported BF16 results include AIME25 89.1 without tools versus 99.2 with tools, GPQA 73.0 without tools, LiveCodeBench v6 68.3, SWE-Bench 38.8 with OpenHands, BFCL v4 53.8 and RULER-100 86.3 at 1M with reasoning off; long-document AA-LCR 35.9 and multilingual MMLU-ProX 59.5 trail Qwen3-30B-A3B-Thinking-2507. Evaluation uses benchmark-specific scaffolds and repeated sampling, not universally single-attempt standalone answers. NVIDIA reports 3.3× throughput versus Qwen3-30B-A3B-Thinking-2507 and 2.2× versus GPT-OSS-20B on one H200 with 8K input/16K output, taking the better vLLM/TRT-LLM configuration; Nemotron and Qwen use FP8, GPT-OSS MXFP4. Selective post-training FP8 quantization retains sensitive attention and preceding Mamba blocks in BF16 and approximately 99% median benchmark accuracy recovery. Original Base-BF16, post-trained BF16 and post-trained FP8 are grouped; later NVFP4, Nano 4B, Omni, Super and Ultra releases are excluded.",
      "paper_url": "https://arxiv.org/abs/2512.20848",
      "links": [
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
        "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8",
        "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16",
        "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
        "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8",
        "https://research.nvidia.com/labs/nemotron/Nemotron-3/",
        "https://huggingface.co/blog/nvidia/nemotron-3-nano-efficient-open-intelligent-models",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/"
      ],
      "page_url": "https://nlp.com.ai/models/nemotron-3-nano-30b-a3b/",
      "record_updated_at": "2026-10-11T08:37:45.434507+00:00"
    },
    {
      "slug": "olmo-31",
      "source_file": "2025-12-12-olmo-3-1.json",
      "release_date": "2025-12-12",
      "date": "December 12, 2025",
      "name": "OLMo 3.1",
      "developer": "Allen Institute for AI (USA)",
      "parameters": "32B Think / 32B Instruct / 7B RL-Zero Math and Code (32,233,522,176 / 7,298,011,136 parameters; dense)",
      "context_window": "65,536 tokens",
      "license": "Apache 2.0 (model weights and core training code); training data have separate ODC-BY and source-specific licenses",
      "key_features": "Primarily English, text-only update to the OLMo 3 model flow: two 32B checkpoints (Think and Instruct) plus refreshed 7B RL-Zero Math and Code. Dense decoder-only Transformers retain untied embeddings, SwiGLU, RMSNorm/QK-Norm, and three 4,096-token sliding-attention layers per full-attention layer; 32B uses 64 layers with 40 query/8 KV heads (GQA), whereas 7B uses 32 layers with 32 query/32 KV heads (MHA). YaRN extends the original 8K context to 65,536 tokens. These models reuse OLMo 3 Base, trained with Dolma 3 pretraining, Dolmino targeted midtraining, and Longmino context extension; this is a post-training update, not a new 6T-token pretraining run. Think retains reasoning SFT and delta-learning DPO, then extends the original OlmoRL run from 750 to 2,300 steps with additional Dolci-Think-RL epochs: Ai2 reports another 21 days on 224 GPUs. OlmoRL incorporates GRPO, DAPO, and Dr GRPO advances, combining verifiable math/code/instruction rewards with model-judged general-chat rewards. Instruct scales the short-response, multi-turn and function-calling recipe to 32B through SFT, DPO and RL; its released final checkpoint suite is limited by a checkpointing issue. RL-Zero learns directly from Base without SFT/DPO; the revised recipe improves stability with 16K training completions and no masking of truncated sequences. Ai2 reports Think scores of 78.1 on AIME 2025, 93.8 on IFEval and 68.1 on IFBench, versus Instruct scores of 57.9, 88.8 and 39.7 respectively. AIME is averaged over 32 sampled answers, not best-of-32; standard post-training evaluations allow up to 32,768 generated tokens. Gains are not uniform: Think AlpacaEval LC falls from 74.2 to 69.1, and knowledge/coding gaps to Qwen remain. Data, code, recipes and checkpoints support inspection, but missing Instruct intermediates and subsequently redacted 7B training PDFs limit exact reproducibility. Original November OLMo 3 checkpoints and RL-Zero Mix are separate releases.",
      "paper_url": "https://arxiv.org/abs/2512.13961",
      "links": [
        "https://allenai.org/blog/olmo3",
        "https://huggingface.co/allenai/Olmo-3.1-32B-Think",
        "https://huggingface.co/allenai/Olmo-3.1-32B-Instruct",
        "https://huggingface.co/allenai/Olmo-3.1-7B-RL-Zero-Math",
        "https://huggingface.co/allenai/Olmo-3.1-7B-RL-Zero-Code",
        "https://github.com/allenai/open-instruct",
        "https://github.com/allenai/OLMo-core"
      ],
      "page_url": "https://nlp.com.ai/models/olmo-31/",
      "record_updated_at": "2026-10-11T08:37:45.432182+00:00"
    },
    {
      "slug": "nomos-1",
      "source_file": "2025-12-09-nomos-1.json",
      "release_date": "2025-12-09",
      "date": "December 9, 2025",
      "name": "Nomos 1",
      "developer": "Nous Research (USA), in collaboration with Hillclimb AI",
      "parameters": "30.532B total; approximately 3.353B active per token (MoE, including input embeddings and output head)",
      "context_window": "262,144 tokens configured; Nomos-specific validated maximum not published",
      "license": "Apache 2.0 (publisher weight declaration; upstream Qwen Apache 2.0 obligations apply); reasoning-harness code is MIT",
      "key_features": "Text-only mathematical reasoning and natural-language proof-writing specialization of Qwen3-30B-A3B-Thinking-2507, trained with Hillclimb AI. The downloadable causal-language-model checkpoint has 48 layers, grouped-query attention with 32 query heads and 4 KV heads, and 128 routed experts with 8 selected per token; it is distinct from the separately released Nomos reasoning harness. Original safetensors store FP32 weights despite a bfloat16 config declaration. The inherited thinking chat template opens <think>; Nous recommends no system prompt and Qwen3 default sampling. Targeted post-training and data quality are credited for the improvement, but the released checkpoint's specific training objectives, SFT/RL stages, dataset size and compute budget are not disclosed; Hillclimb later identifies an olympiad-problem training-corpus preview. With the Nomos harness, the publisher reports 87/120 on Putnam 2025 versus 24/120 for the parent checkpoint under the same conditions, with eight perfect 10/10 answers and one 7/10 answer. These are human blind-graded, multi-attempt system results, not standalone pass@1 or an official competition placing: the published runbooks use three hours for each six-problem section, up to 128 concurrent requests, self-scoring, conclusion consolidation and pairwise tournament selection. The harness uses natural-language model judging, not Lean verification; actual attempt counts, generated-token totals and hardware costs are not reported. Published grader comments identify false lemmas in A6, an incorrect bound in B6, and omitted computational justification in A5. The 262,144-position config is not evidence of Nomos-specific long-context validation, and the tokenizer's 1,010,000-token metadata is not a verified 1M context claim.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/NousResearch/nomos-1",
        "https://github.com/NousResearch/nomos",
        "https://x.com/NousResearch/status/1998536543565127968",
        "https://data.hillclimb.ing/"
      ],
      "page_url": "https://nlp.com.ai/models/nomos-1/",
      "record_updated_at": "2026-10-11T08:37:45.430026+00:00"
    },
    {
      "slug": "devstral-2-devstral-small-2-2512",
      "source_file": "2025-12-09-devstral-2.json",
      "release_date": "2025-12-09",
      "date": "December 9, 2025",
      "name": "Devstral 2 / Devstral Small 2 (2512)",
      "developer": "Mistral AI (France)",
      "parameters": "123B nominal Devstral 2 (125,025,988,608 learned parameters); 24B Devstral Small 2 (24,011,361,280 including vision and connector); dense",
      "context_window": "256K (262,144 tokens) supported for both; Small 2 configuration ceiling: 393,216 tokens",
      "license": "Devstral 2: Modified MIT, retaining attribution/permission notices; no rights under this license when company/employer global consolidated monthly revenue for the preceding month exceeds $20 million (or equivalent currency), including derivatives/modifications/combined works; a discretionary commercial license or Mistral hosted service is an alternative. Exactly $20 million does not trigger the stated exceeds condition. Devstral Small 2: Apache 2.0 publisher grant; standalone LICENSE file absent from the checked repository.",
      "key_features": "Two instruction-tuned, FP8 open-weight coding-agent checkpoints released together for tool-driven codebase exploration, multi-file edits, debugging and software-engineering workflows. Devstral 2 is text-only: an 88-layer dense decoder with 96 query heads, 8 KV heads and untied embeddings. Devstral Small 2 adds image inputs through a 24-layer Pixtral vision encoder and patch-merge connector to a 40-layer dense decoder with 32 query heads and 8 KV heads; its model card identifies Mistral-Small-3.1-24B-Base-2503 as the base and describes Ministral-3-style attention-temperature/RoPE scaling. Both advertise 256K context; the Small checkpoint's larger positional ceiling is not a verified 384K supported window. Publisher-reported SWE-bench Verified scores are 72.2% and 68.0%, SWE-bench Multilingual 61.3% and 55.7%, and Terminal-Bench 2 32.6% and 22.5%, respectively. These are agent-system results, not demonstrated bare-model pass@1: the launch sources do not specify the benchmark scaffold, tool-turn budget, sampling, retries or aggregation. In separate human evaluations scaffolded through Cline, Devstral 2 recorded a 42.8% win rate and 28.6% loss rate against DeepSeek V3.2, while Claude Sonnet 4.5 remained significantly preferred. Mistral claims up to 7x better real-world cost efficiency than Claude Sonnet. Released alongside the Apache-2.0 Mistral Vibe CLI, with file, shell, code-search and version-control tools; integrations include Cline, Kilo Code and Zed. Mistral recommends at least four H100-class GPUs for Devstral 2 and describes Small 2 as locally deployable on consumer hardware; single-GPU/32GB-Mac claims depend on quantization, runtime and context allocation. Both are documented as instruction fine-tunes, and a publisher response confirms Small 2 was trained natively in FP8; the complete release-specific data mixture, training budget and optimization stages are undisclosed. No predecessor training recipe is assumed.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/devstral-2-vibe-cli/",
        "https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512",
        "https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512",
        "https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512/blob/main/LICENSE",
        "https://github.com/mistralai/mistral-vibe"
      ],
      "page_url": "https://nlp.com.ai/models/devstral-2-devstral-small-2-2512/",
      "record_updated_at": "2026-10-11T08:37:45.427542+00:00"
    },
    {
      "slug": "glm-46v-glm-46v-flash",
      "source_file": "2025-12-08-glm-4-6v.json",
      "release_date": "2025-12-08",
      "date": "December 8, 2025",
      "name": "GLM-4.6V / GLM-4.6V-Flash",
      "developer": "Zhipu AI (Z.ai, China)",
      "parameters": "GLM-4.6V: nominal 106B total / 12B active MoE (107.711B stored including vision/projector); Flash: nominal 9B dense (10.293B including vision/projector)",
      "context_window": "128K (131,072 tokens)",
      "license": "MIT (publisher-declared weights); AIMv2 encoder upstream rights require review",
      "key_features": "Same-launch hybrid-reasoning vision-language family: GLM-4.6V and its official FP8 package, plus dense GLM-4.6V-Flash. The updated GLM-V report identifies GLM-4.5-Air and GLM-4-9B-0414 language backbones respectively, with an AIMv2-Huge-initialized vision encoder and visual-to-language adapter. The main decoder has 46 layers, 128 routed experts with eight selected per token and one shared expert, and 96 query / eight KV heads; Flash has 40 dense layers and 32 query / two KV heads. Both accept text, images and video and generate text, with thinking/non-thinking control. Native multimodal function calling is implemented through XML tool calls and image/video-bearing tool responses; URL-referenced visuals enable tool-assisted search, cropping and interleaved illustrated reports, not native image synthesis or built-in execution of arbitrary tools. Documents are understood as rendered page images; file parsing, retrieval and rendering still require the surrounding application. The launch advertises screenshot-to-HTML/CSS/JS replication, visual edits, long-document analysis and approximately 150 document pages, 200 slides or a one-hour video in one pass, dependent on visual tokenization. Training includes multimodal pre-training, long-context continual pre-training, long-CoT supervised fine-tuning and multimodal reinforcement learning with curriculum sampling; the launch additionally describes billion-scale perception/world-knowledge data, synthetic agentic data and tool-invocation rewards. The January 2026 report update explicitly describes a 131,072-token extension and SFT for GLM-4.6V; exact checkpoint-specific training budgets and Flash agentic-stage details are not fully disclosed. Launch evaluation covers over 20 multimodal benchmarks; the retrospective report gives main-model thinking scores of 76.0 MMMU, 85.2 MathVista and 54.9 MMLongBench-Doc, with both gains and regressions versus GLM-4.5V rather than universal superiority on 42 tasks. Agent and coding results depend on external scaffolds and judges; launch-recommended generation uses temperature 0.8, top_p 0.6, top_k 2, repetition penalty 1.1 and a 16K output budget, not a separate native context limit. The publisher notes weaker pure-text QA, overthinking/repetition and counting or identity-recognition errors.",
      "paper_url": "https://arxiv.org/abs/2507.01006v6",
      "links": [
        "https://z.ai/blog/glm-4.6v",
        "https://github.com/zai-org/GLM-V",
        "https://huggingface.co/zai-org/GLM-4.6V",
        "https://huggingface.co/zai-org/GLM-4.6V-FP8",
        "https://huggingface.co/zai-org/GLM-4.6V-Flash",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.6V",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.6V-FP8",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.6V-Flash"
      ],
      "page_url": "https://nlp.com.ai/models/glm-46v-glm-46v-flash/",
      "record_updated_at": "2026-10-11T08:37:45.424990+00:00"
    },
    {
      "slug": "rnj-1-base-and-instruct",
      "source_file": "2025-12-05-rnj-1.json",
      "release_date": "2025-12-05",
      "date": "December 5, 2025",
      "name": "Rnj-1 (Base and Instruct)",
      "developer": "Essential AI (USA)",
      "parameters": "8.31B each (8,310,501,376; dense, tied embeddings)",
      "context_window": "32K (32,768 tokens)",
      "license": "Apache 2.0",
      "key_features": "Essential AI's first open-weight model family, comprising text-only Base and instruction-tuned checkpoints trained from scratch for coding, mathematics and scientific reasoning. Uses a Gemma-3-like dense decoder architecture rather than Gemma-derived weights: 32 layers, 4,096 hidden width, 16,384 MLP width, grouped-query attention with 32 query heads and 8 KV heads, tied embeddings and global attention throughout. Pretraining used 8.4T tokens at 8K context with Muon and a warmup-stable-decay learning-rate schedule; a further 380B-token mid-training stage extended context to 32K with YaRN, followed by 150B tokens of supervised fine-tuning for Instruct. Training included fill-in-the-middle code, program execution and code evolution; no released-model reinforcement-learning stage is documented. Supports multilingual code generation, code infilling and tool calls in <tool_call> tags, with demonstrated agentic coding and profiler-assisted optimization. Publisher reports 20.8% SWE-bench Verified with mini-swe-agent's bash-only scaffold, exceeding Gemini 2.0 Flash and Qwen2.5-Coder 32B Instruct under that framework, and strong BFCL, math and science results among comparable-size models. Reported HE-FIM-Python averages are 82.49% for Base and 86.21% for Instruct. Separate test-time-scaling results use pass@1/2/4/8 from eight generations at temperature 0.2, not single-response guarantees; comparator settings include GPT-OSS-20B at low reasoning effort and Qwen3-8B with thinking disabled. Post-training was deliberately limited to leave room for community specialization. Publisher notes weak factual recall, identity and knowledge-cutoff hallucinations, and a tendency to emit code on non-code prompts. December updates recommend a system prompt and temperature 0–0.2 to mitigate premature truncation; optional 128K YaRN extrapolation requires configuration changes and shows some regressions, so it is not the launch checkpoint's native context.",
      "paper_url": "",
      "links": [
        "https://www.essential.ai/research/rnj-1",
        "https://huggingface.co/EssentialAI/rnj-1",
        "https://huggingface.co/EssentialAI/rnj-1-instruct",
        "https://x.com/essential_ai/status/1997123628765524132",
        "https://x.com/essential_ai/status/1997123638353711411"
      ],
      "page_url": "https://nlp.com.ai/models/rnj-1-base-and-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.422437+00:00"
    },
    {
      "slug": "hermes-43-36b",
      "source_file": "2025-12-03-hermes-4-3.json",
      "release_date": "2025-12-03",
      "date": "December 3, 2025",
      "name": "Hermes 4.3 36B",
      "developer": "Nous Research (USA)",
      "parameters": "36,151,104,512 per checkpoint (36B dense; Psyche and centralized variants)",
      "context_window": "512K (524,288 tokens advertised/configured; centralized training preset uses 131,072-token sequences)",
      "license": "Apache 2.0 declared for Psyche BF16 and official GGUF; centralized derivative license not explicitly stated. Upstream Seed Apache 2.0 notice/redistribution obligations apply; standalone derivative LICENSE files absent.",
      "key_features": "Text-only instruction/reasoning finetunes of ByteDance-Seed/Seed-OSS-36B-Base, not Hermes 4's Llama checkpoints. Same launch offers the production Psyche-trained BF16 model, an independently centrally trained BF16 research artifact, and official GGUF conversions of the Psyche model. Each BF16 checkpoint has 64 dense Transformer layers, 5,120 hidden width, grouped-query attention with 80 query/8 KV heads, untied embeddings and no exported vision or MTP module. Nous calls this its first production model post-trained entirely on Psyche: TP+DisTrO distributes training over the internet, with Solana coordinating consensus and a separate P2P gradient network; the comparison run uses its TorchTitan fork with FSDP+AdamW. The publisher reports 144K training tokens/second across 24 Psyche nodes and a run twice as large as Hermes 4, not foundation pretraining from scratch. Its stated Hermes 4 recipe is supervised instruction/reasoning training using verified synthetic trajectories; Atropos rejection sampling is data generation, not evidence of DPO/GRPO or on-policy RL on the released policy. Exact 4.3 corpus and processed-token totals are not established by the copied Hermes 4 marketing figures. The Seed base already configures 512K, so this is not a verified 128K-to-512K extension of that parent. Llama-style role headers are actually used despite the Seed architecture. The template defaults thinking=False and keep_cots=False; thinking=True selects a reasoning system prompt only when the user has not supplied the first system message. keep_cots controls historical reasoning retention, not a hard decoding budget. Tools are serialized inside <tools>, calls inside <tool_call>, and replies inside <tool_response>; structured JSON and reduced refusal are training goals. Publisher evaluations for the Psyche checkpoint report AIME 2024 71.9% and AIME 2025 69.3% estimated pass@1 from 64 samples, GPQA Diamond 65.5% from eight samples, MATH-500 93.8% from four samples, and IFEval 77.9% instruction-level loose accuracy (65.4% prompt-level strict). This is not best-of-64 or a tool-agent success rate. It exceeds the centralized 36B comparison across the published suite but does not uniformly exceed Hermes 4 70B: SimpleQA fuzzy-match is 6.0% versus 17.9%, and several reasoning/knowledge scores are lower. RefusalBench reports 74.60% non-reasoning and 72.29% reasoning, averaged over five trials; the claimed non-abliterated-model SOTA measures refusal behavior, not universal correctness or safety. Evaluation logs also report 3.4% of samples missing a closing thinking tag, including 18.8% on AIME 2025; exact server-side output limits and tool/retry settings are not fully specified. Quantized GGUF deployment is distinct from the approximately 72.3GB BF16 weight payload and does not guarantee consumer-GPU fit at full 512K context. December 3 is the publisher-declared public release/download-offer date; exact first public unlock remains unresolved. The centralized package lacks an explicit derivative license declaration, and both BF16 packages retain Seed EOS id 2 while the tokenizer's chat EOS is id 155127, requiring serving/stop-token care.",
      "paper_url": "",
      "links": [
        "https://nousresearch.com/introducing-hermes-4-3",
        "https://nousresearch.com/releases",
        "https://huggingface.co/NousResearch/Hermes-4.3-36B",
        "https://huggingface.co/NousResearch/Hermes-4.3-36B-centralized",
        "https://huggingface.co/NousResearch/Hermes-4.3-36B-GGUF",
        "https://huggingface.co/datasets/NousResearch/eval-Hermes-4.3-36B",
        "https://huggingface.co/datasets/NousResearch/eval-Hermes-4.3-36B-centralized",
        "https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base",
        "https://arxiv.org/abs/2508.18255v1"
      ],
      "page_url": "https://nlp.com.ai/models/hermes-43-36b/",
      "record_updated_at": "2026-10-11T08:37:45.420057+00:00"
    },
    {
      "slug": "mistral-large-3",
      "source_file": "2025-12-02-mistral-large-3.json",
      "release_date": "2025-12-02",
      "date": "December 2, 2025",
      "name": "Mistral Large 3",
      "developer": "Mistral AI (France)",
      "parameters": "675B total / 41B active (publisher nominal; granular MoE with a 2.5B vision encoder)",
      "context_window": "256K (262,144 tokens supported; 294,912 positional ceiling in release configs)",
      "license": "Apache 2.0 (explicit weight-license grant in release model cards; third-party rights must be respected)",
      "key_features": "Multimodal, multilingual granular mixture-of-experts family released as a BF16 Base checkpoint and an instruction-tuned checkpoint in FP8 and NVFP4; separate from the dense Ministral 3 family. Its 61-layer language model uses multi-head latent attention, three initial dense layers, and 128 routed experts per MoE layer with four selected per token plus one shared expert; a 48-layer vision encoder supports image understanding. The complete Base export contains 675,991,176,960 parameters: 673,421,089,792 in the language model, 2,495,702,144 in the vision encoder, and 74,385,024 in the multimodal connector and normalization layers. The publisher describes approximately 39B active language parameters and 41B active overall; these nominal figures use an unspecified counting convention. Mistral reports training from scratch on 3,000 H200 GPUs, followed by instruction fine-tuning; corpus size, token budget, and detailed SFT/preference/RL stages are not disclosed. Instruct supports system prompts, native function calling, JSON output, coding, long-document understanding, and multilingual enterprise assistants. At launch Mistral reported #2 among OSS non-reasoning models and #6 among OSS models overall on LMArena; its published chart gives 1418 +/- 11 Elo. Third-party human evaluations reported 53%/55% general-prompt and 57%/60% multilingual-prompt wins against DeepSeek V3.1/Kimi K2, respectively, without publishing sample counts or generation budgets. FP8 is advertised for one eight-H200/B200 node, and the llm-compressor NVFP4 export for one eight-H100/A100 node. An Instruct BF16 export is also publicly downloadable, but its exact same-day availability remains under review. The EAGLE checkpoint is an auxiliary speculative-decoding draft, not an additional 675B chat model. Official limitations include weaker strict reasoning than dedicated reasoning models, weaker multimodal performance than vision-first models, and demanding deployment requirements.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/mistral-3",
        "https://huggingface.co/collections/mistralai/mistral-large-3",
        "https://huggingface.co/mistralai/Mistral-Large-3-675B-Base-2512",
        "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512",
        "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4",
        "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512-BF16",
        "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512-Eagle"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-large-3/",
      "record_updated_at": "2026-10-11T08:37:45.417469+00:00"
    },
    {
      "slug": "ministral-3",
      "source_file": "2025-12-02-ministral-3.json",
      "release_date": "2025-12-02",
      "date": "December 2, 2025",
      "name": "Ministral 3",
      "developer": "Mistral AI (France)",
      "parameters": "3B / 8B / 14B dense family; approximately 3.849B / 8.918B / 13.945B total learned parameters including vision and connectors",
      "context_window": "Base/Instruct: 262,144 tokens; Reasoning: 128K reported support, despite 262,144-token launch configs",
      "license": "Apache 2.0",
      "key_features": "Nine multimodal checkpoints: 3B, 8B and 14B, each released as Base, Instruct and Reasoning. Dense decoder-only transformers with 26/34/40 layers, grouped-query attention (32 query heads, 8 KV heads), SwiGLU and YaRN long-context extension; all accept text and images and generate text, using a shared Pixtral-style vision encoder. The January 2026 Ministral 3 report describes the released family as descendants of Mistral Small 3.1 Base through Cascade Distillation: successive pruning and continued pretraining with teacher-logit distillation on text and interleaved image/text data, followed by long-context extension from 16K to 256K. The vision encoder is frozen and each size learns a new connector; the 3B decoder ties input/output embeddings. Instruct models undergo FP8-aware supervised logit distillation from Mistral Medium 3 followed by online DPO with reward-model-ranked on-policy responses and tool execution. Reasoning models start from Base, not Instruct, and use chain-of-thought SFT, STEM and general-task GRPO, then online DPO; the 3B reasoning SFT additionally distills Magistral Small 1.2. Base/Reasoning launch weights are BF16, while Instruct weights use mixed FP8/BF16; official 14B Instruct BF16 and GGUF downloads were also offered on launch day. Mistral reports 85.0% AIME 2025 and 71.2% GPQA Diamond for 14B Reasoning, with the report labeling reasoning evaluations pass@16 and LiveCodeBench pass@5, not independently verified single-attempt results; launch sampling, tools and retry details are incomplete. Multilingual, image understanding, function calling and JSON output support target local assistants and edge deployment. Instruct favors concise responses while Reasoning can produce longer reasoning traces. Small-device and single-GPU deployment depend on precision, hardware and context budget: fitting weights is not a guarantee of full-context inference. Reasoning support is stated as 128K in the report, conflicting with launch cards/configs advertising 256K; the 80K generation length used during reasoning RL is a training setting, not a verified universal output limit.",
      "paper_url": "https://arxiv.org/abs/2601.08584v1",
      "links": [
        "https://mistral.ai/news/mistral-3",
        "https://huggingface.co/collections/mistralai/ministral-3",
        "https://huggingface.co/mistralai/Ministral-3-3B-Base-2512",
        "https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512",
        "https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512",
        "https://huggingface.co/mistralai/Ministral-3-8B-Base-2512",
        "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512",
        "https://huggingface.co/mistralai/Ministral-3-8B-Reasoning-2512",
        "https://huggingface.co/mistralai/Ministral-3-14B-Base-2512",
        "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512",
        "https://huggingface.co/mistralai/Ministral-3-14B-Reasoning-2512",
        "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512-BF16",
        "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512-GGUF",
        "https://huggingface.co/collections/mistralai/ministral-3-additional-checkpoints"
      ],
      "page_url": "https://nlp.com.ai/models/ministral-3/",
      "record_updated_at": "2026-10-11T08:37:45.414936+00:00"
    },
    {
      "slug": "trinity-nano-preview-mini",
      "source_file": "2025-12-01-trinity.json",
      "release_date": "2025-12-01",
      "date": "December 1, 2025",
      "name": "Trinity Nano Preview / Mini",
      "developer": "Arcee AI (USA)",
      "parameters": "Nano Preview: 6.120B total / ~1B active; Mini: 26.124B total / ~3B active (MoE)",
      "context_window": "131,072 tokens (128K)",
      "license": "Apache 2.0 at release; OpenMDW-1.1 since May 28, 2026",
      "key_features": "Two text-only, decoder-only AFMoE models trained from scratch in the U.S.: experimental, personality-forward Nano Preview for low-resource chat and instruction following, and fully post-trained Mini for reasoning, tool calling and multi-step agent workflows. Nano has 56 layers and Mini 32; both use 128 routed experts with 8 selected per token plus one shared expert, two initial dense layers, gated grouped-query attention with QK normalization, and a 3:1 pattern of 2,048-token local RoPE attention and global NoPE attention. Each was pretrained on 10T Datology-curated tokens in 7T/1.8T/1.2T phases with increasing math and code concentration, using Muon for hidden layers, AdamW for embeddings/output, modified TorchTitan and HSDP on up to 512 H200 GPUs. Nano underwent 256K-sequence context extension for 128K inference; Mini was extended at 128K. The released Nano Preview is not a thinking model and may be unstable in edge cases. Mini is reasoning-tuned, but the released Nano/Mini post-training objectives and budgets are not fully disclosed; the later family report's SFT/RL section concerns Trinity Large, not these checkpoints. Arcee reports Mini scores of 84.95 on zero-shot MMLU, 92.10 on Math-500, 58.55 on GPQA-Diamond and 59.67 on BFCL V3, alongside 63.49 on MUSR and 8.90 on SimpleQA; sampling, tool scaffolds, retry budgets and aggregation are unspecified. Base, pre-anneal and later Large/precision variants are not silently grouped with this two-model launch.",
      "paper_url": "https://arxiv.org/abs/2602.17004v1",
      "links": [
        "https://www.arcee.ai/blog/the-trinity-manifesto",
        "https://x.com/arcee_ai/status/1995600354374025395",
        "https://huggingface.co/arcee-ai/Trinity-Nano-Preview",
        "https://huggingface.co/arcee-ai/Trinity-Mini",
        "https://github.com/arcee-ai/trinity-large-tech-report"
      ],
      "page_url": "https://nlp.com.ai/models/trinity-nano-preview-mini/",
      "record_updated_at": "2026-10-11T08:37:45.412559+00:00"
    },
    {
      "slug": "deepseek-v32-v32-and-v32-speciale",
      "source_file": "2025-12-01-deepseek-v3-2.json",
      "release_date": "2025-12-01",
      "date": "December 1, 2025",
      "name": "DeepSeek-V3.2 (V3.2 and V3.2-Speciale)",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B nominal total / 37B nominal active per token (each MoE); 671.878B main-model tensors plus 13.477B packaged MTP tensors",
      "context_window": "128K tokens supported; 163,840 configured positional ceiling",
      "license": "MIT",
      "key_features": "Two text-only post-trained checkpoints sharing a 61-layer DeepSeekMoE/MLA architecture with 256 routed experts, eight selected per token and one shared expert. DeepSeek Sparse Attention uses a 64-head, 128-dimensional lightning indexer selecting up to 2,048 prior tokens. Released weights mix FP8 E4M3 and BF16, with UE8M0-format block scaling. The shared sparse base underwent 2.1B-token indexer warm-up and 943.7B-token sparse continued pretraining; these are inherited base-training stages, not an additional December run. Standard V3.2 uses specialist-generated thinking/non-thinking distillation data followed by thousands of mixed GRPO steps for reasoning, agents and human alignment, with length and language-consistency rewards. Agent synthesis includes 1,827 general-agent environments and over 85,000 prompts across four agent domains. Speciale instead trains exclusively on reasoning data with reduced length penalties and DeepSeekMath-V2 proof data/rewards; it consumes more tokens and does not support tool calling. Standard V3.2 is DeepSeek's first model to combine thinking with tools, while also supporting non-thinking tool use. Developer-reported pass@1 is 93.1% AIME 2025 and 92.5% HMMT February for standard, versus 96.0% and 99.2% for Speciale. Standard achieves 73.1% SWE-bench Verified with an internal agent framework and 46.4% Terminal Bench 2.0 with Claude Code; BrowseComp rises from 51.4% to 67.6% with context management. Speciale reaches gold-level competition scores (IMO 35/42, CMO 102/126, IOI 492/600, ICPC 10/12) using filtered multi-candidate submissions or generate-verify-refine workflows, not standalone one-shot medals. The report identifies knowledge breadth, token efficiency and long-trajectory context overflow as remaining weaknesses. Launch packaging supplies Python message encoding rather than a Jinja chat template; its output parser needs production error handling.",
      "paper_url": "https://arxiv.org/abs/2512.02556v1",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.2",
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Speciale",
        "https://www.deepseek.com/zh/news/deepseek-v3-2",
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/c69397ecfd1fd142e90e3fbad51f4c7e40b9f3d3/assets/paper.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v32-v32-and-v32-speciale/",
      "record_updated_at": "2026-10-11T08:37:45.409992+00:00"
    },
    {
      "slug": "deepseekmath-v2",
      "source_file": "2025-11-27-deepseek-math-v2.json",
      "release_date": "2025-11-27",
      "date": "November 27, 2025",
      "name": "DeepSeekMath-V2",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B-class MoE (37B nominal active backbone); 671.878B main non-scale tensors plus 13.477B exported MTP tensors; approximately 685.4B stored tensor elements including scales and routing buffers",
      "context_window": "128K-token evaluation limit; 163,840 configured positions",
      "license": "Apache-2.0 for the published checkpoint and code; early November 27 HF revisions carried MIT before the same-day Apache replacement, and upstream MIT notices remain relevant",
      "key_features": "Open-weight text-generation checkpoint for natural-language mathematical proofs, built on DeepSeek-V3.2-Exp-Base. The released final proof generator also performs proof verification; generator, verifier and meta-verifier are training roles, not a shipped two-model architecture. Its 61-layer MLA/MoE backbone selects eight of 256 routed experts plus one shared expert, and the released config includes DeepSeek Sparse Attention with a 64-head, 128-dimensional lightning indexer selecting 2,048 key-value entries. Weights use mixed FP8/BF16 with UE8M0 scales and include an auxiliary MTP block. Training alternates GRPO optimization of proof verification and generation: the initial verifier starts from a mathematics/code reasoning SFT checkpoint, learns expert proof scores of 0, 0.5 or 1, and uses meta-verification to penalize hallucinated defects. Each generator is initialized from the verifier; subsequent iterations consolidate verification and generation through rejection fine-tuning. Generator rewards combine proof quality (weight 0.76) with faithful self-assessment (0.24), encouraging a rigorous Solution followed by Self Evaluation and correction of identified errors. Scaled verification automatically labels difficult proofs, replacing human annotation in the last two training iterations. DeepSeek reports gold-level, not officially awarded medal, performance on IMO 2025 (five of six problems, 83.3% of points) and CMO 2024 (four fully solved plus partial credit, 73.8%), and 118/120 on Putnam 2024. These are high-compute search results: 64 initial proofs, 64 verification analyses per proof, refinement of the best 64 using eight analyses each, and up to 16 iterations; the same final generator supplies generation and verification, with mathematical experts grading selected proofs. Heavy-mode IMO-ProofBench expert evaluations report 99.0% Basic and 61.9% Advanced, not one-shot pass rates. The hardest IMO-level problems remain challenging, and natural-language self-verification is not formal proof certification. The report uses a 128K-token per-attempt limit; the larger positional config is not independently validated context support. November 27 is corroborated availability-by, with exact first public weight access unresolved. This card covers only the original final checkpoint, not separate unreleased training-stage verifiers, earlier 7B DeepSeekMath models or a hosted API alias.",
      "paper_url": "https://arxiv.org/abs/2511.22570",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-Math-V2",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-Math-V2",
        "https://github.com/deepseek-ai/DeepSeek-Math-V2"
      ],
      "page_url": "https://nlp.com.ai/models/deepseekmath-v2/",
      "record_updated_at": "2026-10-11T08:37:45.407249+00:00"
    },
    {
      "slug": "intellect-3",
      "source_file": "2025-11-26-intellect-3.json",
      "release_date": "2025-11-26",
      "date": "November 26, 2025",
      "name": "INTELLECT-3",
      "developer": "Prime Intellect (USA)",
      "parameters": "106B total / 12B active (MoE; BF16 package contains 106,852,245,504 BF16 parameter elements plus 5,760 FP32 routing-correction elements)",
      "context_window": "131,072 tokens (configured positional ceiling; tokenizer model_max_length: 128,000)",
      "license": "MIT (official weight metadata and release declaration; standalone LICENSE absent from both HF precision packages; upstream GLM-4.5-Air-Base is also declared MIT, with notice-retention obligations)",
      "key_features": "Text-only, always-reasoning model post-trained from GLM-4.5-Air-Base, not MiniMax-M2 or a new model pretrained from scratch. The released decoder has 46 layers, hidden width 4,096, grouped-query attention with 96 query and 8 KV heads, one initial dense layer followed by MoE layers with 128 routed experts, eight selected per token, and one shared expert; vocabulary size is 151,552 with untied embeddings. Although config declares one next-token prediction layer, neither BF16 nor FP8 weight index exports an additional MTP module. General chat/reasoning SFT uses DeepSeek-R1-0528 traces from NVIDIA Nemotron and AM datasets, followed by agentic SFT with SWE-Swiss, Toucan and environment-generated trajectories, then asynchronous RL across math, code, science, logic, deep research and software engineering. The report specifies IcePop masked token-level importance sampling with group-mean reward advantages, Muon optimization, continuous batching, in-flight policy updates and up to eight off-policy steps. SFT and RL, including ablations, used a 512-H200 cluster over approximately two months; the described RL allocation is 128 trainer GPUs plus 352 rollout GPUs. SFT sequence lengths of 65K and 98K and the 65,536-token RL cap are training settings, not independently validated inference-context or reasoning-output limits. Publisher evaluations report AIME 2024 90.8% and AIME 2025 88.0% as Avg@32 estimates of pass@1, GPQA Diamond 74.4% as Avg@4, LiveCodeBench v6 69.3% as Avg@2, text-only tool-free HLE 14.6%, and MMLU-Pro 81.9%. These are not majority-vote or best-of-N scores. This candidate covers the original BF16 checkpoint; the official mixed FP8 package is an eligible related precision variant, but exact same-day public download availability remains unconfirmed. Publisher serving guidance is two H200s for BF16 or one for FP8. November 26 is the publisher's dated explicit weight-release offer, not a verified first-public-unlock timestamp. Supported context beyond the configured/tokenizer limits and an end-to-end training cost were not independently established.",
      "paper_url": "https://arxiv.org/abs/2512.16144",
      "links": [
        "https://huggingface.co/PrimeIntellect/INTELLECT-3",
        "https://huggingface.co/PrimeIntellect/INTELLECT-3-FP8",
        "https://www.primeintellect.ai/blog/intellect-3",
        "https://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf",
        "https://github.com/PrimeIntellect-ai/prime-rl",
        "https://github.com/PrimeIntellect-ai/verifiers"
      ],
      "page_url": "https://nlp.com.ai/models/intellect-3/",
      "record_updated_at": "2026-10-11T08:37:45.404457+00:00"
    },
    {
      "slug": "llada20-mini-and-flash",
      "source_file": "2025-11-25-llada-2-0.json",
      "release_date": "2025-11-25",
      "date": "November 25, 2025",
      "name": "LLaDA2.0 (mini and flash)",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "mini: 16B nominal non-embedding / 1.4B active (16,255,643,392 stored); flash: 100B nominal non-embedding / 6.1B active (102,889,705,216 stored); MoE",
      "context_window": "32,768 tokens (both variants)",
      "license": "Apache 2.0 (both checkpoints; retain applicable MIT notices from Ling ancestry)",
      "key_features": "Instruction-tuned, text-only block-diffusion MoE models converted from the Ling-mini-2.0 and Ling-flash-2.0 autoregressive model family, rather than trained from scratch like the original LLaDA. Tokens are iteratively unmasked in parallel with bidirectional attention inside each block; blocks are generated sequentially, conditioned on preceding blocks. The dedicated LLaDA2.0 report describes continual pretraining with a Warmup-Stable-Decay block-size curriculum: increase from 1 through 4, 32 and 64 to full-sequence 4,096-token diffusion, train at that scale, then reduce to compact blocks such as 32. Document-level attention prevents cross-document leakage, and top-k checkpoint averaging consolidates the converted model. Post-training uses block-diffusion SFT with bounded masking rates and complementary masks, followed by diffusion-ELBO-based DPO on 1.5 million preference pairs; reinforcement-learning enhancement was announced as future work. Mini has 20 layers and flash 32; both use 256 routed experts with eight selected per token plus one shared expert, RoPE and a 157,184-token vocabulary. The release emphasizes coding, mathematical reasoning and tool calling; official model cards report HumanEval 86.59/94.51 and AIME 2025 36.67/60.00 for mini/flash. Recommended decoding uses temperature 0, 32-token blocks and up to 32 refinement steps per block, with confidence threshold 0.95. Native context is 32K; the report tests 64K with inference-time YaRN scaling at reduced accuracy, not a native 64K release. Optimized block decoding can reuse KV cache in dInfer/SGLang; speed depends on decoder and serving setup. Earlier preview checkpoints and later confidence-aware-parallel (CAP) checkpoints are separate releases and are not included here.",
      "paper_url": "https://arxiv.org/abs/2512.15745v1",
      "links": [
        "https://huggingface.co/inclusionAI/LLaDA2.0-mini",
        "https://huggingface.co/inclusionAI/LLaDA2.0-flash",
        "https://modelscope.cn/models/inclusionAI/LLaDA2.0-mini",
        "https://modelscope.cn/models/inclusionAI/LLaDA2.0-flash",
        "https://github.com/inclusionAI/LLaDA2.X"
      ],
      "page_url": "https://nlp.com.ai/models/llada20-mini-and-flash/",
      "record_updated_at": "2026-10-11T08:37:45.401534+00:00"
    },
    {
      "slug": "nanbeige4-3b-thinking-2511",
      "source_file": "2025-11-21-nanbeige4-3b-thinking-2511.json",
      "release_date": "2025-11-21",
      "date": "November 21, 2025",
      "name": "Nanbeige4-3B-Thinking-2511",
      "developer": "Nanbeige LLM Lab / BOSS Zhipin (China)",
      "parameters": "3B nominal; 3,933,637,120 total (dense, untied embeddings)",
      "context_window": "65,536 tokens",
      "license": "Apache 2.0 (model-card declaration; no standalone LICENSE file in checkpoint repository)",
      "key_features": "English/Chinese text-only reasoning and chat model, an enhanced iteration of Nanbeige4-3B-Thinking-2510 built on Nanbeige4-3B-Base rather than a Qwen derivative. Uses a Llama-compatible decoder with 32 layers, hidden size 2,560, 20 attention heads and four key/value heads; the approximately 3.08B non-embedding parameters become 3.93B including separate input embeddings and output head. The Nanbeige4 technical report describes 23T-token pretraining with hybrid quality filtering and a Fine-Grained Warmup-Stable-Decay curriculum, extending context to 64K through adjusted RoPE base frequency. Post-training combines approximately 30 million reasoning QA samples at 32K, broader 64K SFT including conversation, writing and tool trajectories, solution refinement with chain-of-thought reconstruction, Dual-Level Preference Distillation from Nanbeige3.5-Pro using token-level distillation plus sequence-level DPO, and multi-stage on-policy GRPO for STEM, practical coding and human-preference alignment. Native function calling is supported by trained tool-use data and a chat template with tool-call/tool-response markup. For the 2511 checkpoint, the report gives AIME 2024 90.4 and AIME 2025 85.6 (avg@8), GPQA-Diamond 82.2, BFCL-V4 53.8 and Arena-Hard V2 60.0; evaluation uses temperature 0.6, top-p 0.95 and a 64K generation ceiling. These are benchmark-specific comparisons, not universal superiority: Fullstack Bench 48.0 trails Qwen3-32B's 58.2, while the model card reports historical WritingBench #11 and EQBench3 #15 rankings. The release is a single Thinking checkpoint, not a separate non-thinking variant; earlier Base/2510 and later Nanbeige4.1 models are separate releases. The publisher warns that harmful or biased outputs remain possible.",
      "paper_url": "https://arxiv.org/abs/2512.06266",
      "links": [
        "https://huggingface.co/Nanbeige/Nanbeige4-3B-Thinking-2511",
        "https://huggingface.co/Nanbeige",
        "https://arxiv.org/abs/2512.06266"
      ],
      "page_url": "https://nlp.com.ai/models/nanbeige4-3b-thinking-2511/",
      "record_updated_at": "2026-10-11T08:37:45.398471+00:00"
    },
    {
      "slug": "olmo-3",
      "source_file": "2025-11-20-olmo-3.json",
      "release_date": "2025-11-20",
      "date": "November 20, 2025",
      "name": "OLMo 3",
      "developer": "Allen Institute for AI (USA)",
      "parameters": "7B / 32B dense (7,298,011,136 / 32,233,522,176 parameters)",
      "context_window": "65,536 tokens",
      "license": "Apache 2.0 (model weights and core training code); training data have separate ODC-BY and source-specific licenses",
      "key_features": "Fully open, primarily English text-generation family: 7B and 32B Base and Think, 7B Instruct, and four experimental 7B RL-Zero pathways (Math, Code, IF, General), with intermediate training checkpoints. Dense decoder-only Transformers use SwiGLU, RMSNorm/QK-Norm, untied embeddings, and three 4,096-token sliding-attention layers per full-attention layer; 7B has 32 layers with 32 query/32 KV heads, while 32B has 64 layers with 40 query/8 KV heads (GQA). YaRN extends the original 8,192-token window to 65,536. Pretraining uses Dolma 3 Mix (5.93T tokens for 7B; 5.50T for 32B) selected from a 9.31T-token pool of web, science PDFs, code, math, and encyclopedic text, followed by Dolmino targeted midtraining and Longmino long-context extension. Think and Instruct use Dolci supervised fine-tuning, DPO preference tuning, and OlmoRL reinforcement learning with verifiable rewards, incorporating GRPO, DAPO, and Dr GRPO advances; RL-Zero studies reinforcement learning directly from Base without the SFT/DPO stages. Think produces explicit reasoning traces for math and coding; Instruct favors shorter responses, multi-turn chat, instruction following, and function calling. Ai2 reports original Think 32B scores of 96.1 on MATH, 76.8 on AIME 2024, and 89.0 on IFEval under its evaluation protocols, while noting remaining gaps to Qwen 3 in some tasks. Training data, recipes, code, and intermediate checkpoints support inspection and customization, but do not guarantee exact reproducibility; some 7B training PDFs have subsequently been redacted. Later OLMo 3.1 checkpoints and RL-Zero Mix are not part of this launch grouping.",
      "paper_url": "https://arxiv.org/abs/2512.13961",
      "links": [
        "https://allenai.org/blog/olmo3",
        "https://huggingface.co/collections/allenai/olmo-3-68e80f043cc0d3c867e7efc6",
        "https://huggingface.co/allenai/Olmo-3-1025-7B",
        "https://huggingface.co/allenai/Olmo-3-1125-32B",
        "https://huggingface.co/allenai/Olmo-3-7B-Think",
        "https://huggingface.co/allenai/Olmo-3-32B-Think",
        "https://huggingface.co/allenai/Olmo-3-7B-Instruct",
        "https://huggingface.co/allenai/Olmo-3-7B-RL-Zero-General",
        "https://huggingface.co/allenai/Olmo-3-7B-RL-Zero-IF",
        "https://huggingface.co/allenai/Olmo-3-7B-RL-Zero-Math",
        "https://huggingface.co/allenai/Olmo-3-7B-RL-Zero-Code",
        "https://github.com/allenai/OLMo-core",
        "https://github.com/allenai/open-instruct"
      ],
      "page_url": "https://nlp.com.ai/models/olmo-3/",
      "record_updated_at": "2026-10-11T08:37:45.395844+00:00"
    },
    {
      "slug": "baguettotron",
      "source_file": "2025-11-12-baguettotron.json",
      "release_date": "2025-11-12",
      "date": "November 12, 2025",
      "name": "Baguettotron",
      "developer": "Pleias (France)",
      "parameters": "321M (320,956,992 total; dense, tied embeddings)",
      "context_window": "4,096 tokens (configured; retrospective report documents 4K context-extension training)",
      "license": "Apache 2.0 (publisher model-card declaration; no standalone LICENSE bundled)",
      "key_features": "Text-only generalist small reasoning model with an unusually deep, narrow Llama decoder: 80 layers, hidden width 576, SwiGLU intermediate width 1,536, grouped-query attention with 9 query heads and 3 KV heads, and a 65,536-token multilingual vocabulary. Trained end-to-end on approximately 200B processed tokens of SYNTH using Nanotron on 16 H100s at Jean Zay; the synthetic corpus supplies instructions, reasoning traces and answers together rather than relying on a conventional web-pretraining followed by instruction-tuning pipeline. SYNTH reports approximately 75B corpus tokens, distinct from the model training-token budget, and amplifies Wikipedia/Wikibooks seeds into encyclopedic QA, source-grounded RAG, arithmetic, editing, extraction and constrained creative writing; code was deliberately excluded. English plus French, German, Italian, Spanish and Polish are supported, with weaker claimed coverage of Dutch and Latin; reasoning traces are English-only. ChatML-style prompting defaults to <think>; manually substituting </think> suppresses thinking but the publisher reports lower performance, especially factual recall. Compact logical, confidence and verification markers structure reasoning; simulated entropy markers do not change decoder temperature. RAG uses supplied <source_n> passages and <ref> quotations, not an autonomous retrieval executor. The launch table reports MMLU 40, GSM8K 39 and HotPotQA 51, with launch evaluation protocol unspecified; these are publisher results, not universal best-in-class performance. Limitations include English/Wikipedia-centered seed coverage, no code-training claim, and launch tokenizer/stop-token inconsistencies requiring runtime care. The September 2026 paper retrospectively documents the original 321M architecture and training; its later 600M, MoE and tool-calling adaptations are not this checkpoint. GRPO is suggested as future fine-tuning, not established training of the released model. Original BF16 weights only; later official GGUF packaging is a separate release-date consideration.",
      "paper_url": "https://arxiv.org/abs/2609.37891",
      "links": [
        "https://huggingface.co/PleIAs/Baguettotron",
        "https://pleias.fr/blog/blogsynth-the-new-data-frontier",
        "https://huggingface.co/datasets/PleIAs/SYNTH"
      ],
      "page_url": "https://nlp.com.ai/models/baguettotron/",
      "record_updated_at": "2026-10-11T08:37:45.393188+00:00"
    },
    {
      "slug": "kimi-k2-thinking",
      "source_file": "2025-11-06-kimi-k2-thinking.json",
      "release_date": "2025-11-06",
      "date": "November 6, 2025",
      "name": "Kimi K2 Thinking",
      "developer": "Moonshot AI (China)",
      "parameters": "1T total / 32B active (MoE)",
      "context_window": "262,144 tokens (256K)",
      "license": "Modified MIT; retain notices and prominently display Kimi K2 in the UI of a commercial product/service exceeding 100 million monthly active users OR US$20 million monthly revenue",
      "key_features": "Text-only thinking-agent checkpoint built from Kimi K2, trained to interleave step-by-step reasoning with tool calls for research, coding and writing. Its MLA/SwiGLU MoE has 61 layers, including one dense layer, 64 attention heads, 384 routed experts with eight selected per token, and one shared expert. This release is the original Kimi-K2-Thinking checkpoint, not the earlier Base/Instruct/0905 or Kimi Linear releases; hosted Thinking Turbo is not an additional verified weight variant. Post-training uses quantization-aware training with INT4 weight-only routed-expert quantization, symmetric groups of 32 and compressed-tensors packaging; attention, shared experts and other excluded components remain higher precision. Moonshot reports roughly 2x faster generation in low-latency inference and reduced GPU-memory use without benchmark-quality loss; these are publisher claims, not universal hardware guarantees. All published benchmark scores use INT4. Reported results include HLE text-only 23.9% without tools and 44.9% with search, browsing and code-interpreter tools, BrowseComp 60.2%, SWE-bench Verified 71.3% and Terminal-Bench 47.1% with JSON simulated tools. HLE tool evaluation permits 120 steps with 48K reasoning tokens per step; agentic search permits 300 steps with 24K per step and hides earlier tool outputs when context overflows. Coding scores average five independent runs: SWE tasks use a modified SWE-agent-derived harness, while Terminal-Bench uses Terminus-2; unspecified per-task retries must not be read as best-of-five or verified single-attempt pass@1. The claimed 200–300 sequential tool calls describe long-horizon capability, not a guaranteed successful workflow or the HLE evaluation budget. Heavy-mode scores use eight parallel trajectories followed by reflective aggregation, not separate weights. Local use requires compatible reasoning/tool parsers and externally executed tools; the launch guide gives an eight-H200 INT4/256K deployment example, and the hosted chat product restricts tools and turns, so it may not reproduce benchmark scores. The inherited K2 foundation report describes MuonClip pretraining on 15.5T tokens, but Thinking-specific data volume, full optimization recipe and training cost are not disclosed in the checked release sources.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-K2-Thinking",
        "https://modelscope.cn/models/moonshotai/Kimi-K2-Thinking",
        "https://moonshotai.github.io/Kimi-K2/thinking.html",
        "https://github.com/MoonshotAI/Kimi-K2",
        "https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main/docs/deploy_guidance.md",
        "https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main/LICENSE",
        "https://arxiv.org/abs/2507.20534v1"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k2-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.390694+00:00"
    },
    {
      "slug": "kimi-linear-48b-a3b-instruct",
      "source_file": "2025-10-30-kimi-linear-48b-a3b-instruct.json",
      "release_date": "2025-10-30",
      "date": "October 30, 2025",
      "name": "Kimi Linear 48B A3B Instruct",
      "developer": "Moonshot AI (China)",
      "parameters": "48B total / 3B active per token (nominal MoE); 49,122,681,728 checkpoint parameters",
      "context_window": "1,048,576 tokens (1M)",
      "license": "MIT (explicit Base/Instruct checkpoint declarations; copyright/permission notices must be preserved; standalone HF LICENSE files absent)",
      "key_features": "Text-generation checkpoint using a layerwise hybrid of Kimi Delta Attention (KDA) and full Multi-Head Latent Attention (MLA), not Kimi K2 or a parallel concatenated attention branch. KDA extends Gated DeltaNet with channel-wise decay gating and a hardware-efficient chunkwise delta-rule algorithm using specialized diagonal-plus-low-rank transitions. The advertised 3:1 KDA/MLA pattern is implemented as 20 KDA and 7 MLA layers across 27 decoder layers; MLA uses NoPE. The first feed-forward layer is dense; subsequent MoE layers select 8 of 256 routed experts plus 1 shared expert. Released weights are predominantly BF16 with FP32 KDA gate parameters and no exported MTP layers. Pretrained on 5.7T tokens sampled from the K2 corpus with MuonClip, a WSD schedule and initial 4,096-token sequences, followed by annealing and long-context activation; both model card and launch config support 1M context without a YaRN override. The report describes multi-stage general/reasoning SFT and math/code/STEM RL with concurrent PTX/SFT, truncated importance sampling and adaptive KL/update scheduling; precise final-checkpoint stage budgets are not published. The shipped chat template has no thinking-mode switch, although reasoning benchmarks and RL training are reported. Publisher-reported Instruct results include RULER 94.8 at 1M, MMLU-Pro 72.7 and LiveCodeBench v6 pass@1 45.7. Controlled 1.4T-token same-backbone comparisons favor Kimi Linear on average, but it trails full MLA on LongBench V2, Frames and commit-level Long Code Arena; these ablations are not the released 5.7T checkpoint. Up to 75% less KV cache and approximately 6x faster decoding at 1M are publisher-specific comparison claims, not universal hardware guarantees: the report distinguishes larger-batch headline TPOT from batch-one tests and does not disclose complete hardware/precision settings in v1. An official eligible Base checkpoint also exists, but is excluded from this provisional Instruct-only card until its first public-download date and same-day grouping are resolved.",
      "paper_url": "https://arxiv.org/abs/2510.26692",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct",
        "https://modelscope.cn/models/moonshotai/Kimi-Linear-48B-A3B-Instruct",
        "https://github.com/MoonshotAI/Kimi-Linear"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-linear-48b-a3b-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.387747+00:00"
    },
    {
      "slug": "minimax-m2",
      "source_file": "2025-10-27-minimax-m2.json",
      "release_date": "2025-10-27",
      "date": "October 27, 2025",
      "name": "MiniMax-M2",
      "developer": "MiniMax AI (China)",
      "parameters": "230B total / 10B active (MoE; 228.690B learned parameters in the released main checkpoint)",
      "context_window": "196,608 tokens (192K native; approximately 200K)",
      "license": "MIT at the October 27 launch; changed November 13, 2025 to Modified MIT with UI attribution for commercial products/services exceeding 100M monthly active users OR US$30M annual recurring revenue",
      "key_features": "Original text-only coding and agentic reasoning checkpoint, released in mixed FP8 precision. Its 62-layer decoder-only Transformer uses full softmax grouped-query attention throughout (48 query heads, 8 KV heads), QK normalization and partial RoPE, not MiniMax-M1's hybrid Lightning Attention. Each MoE layer routes tokens to 8 of 256 fine-grained experts with sigmoid gating and no shared expert. Supports multi-file editing, code-run-test-repair workflows and structured XML tool calls; shell, browser, retrieval and Python execution require an external agent runtime. Interleaved reasoning uses <think>...</think>; MiniMax recommends preserving prior thinking in conversation history. The retrospective May 2026 M2-series report describes 29.2T pretraining tokens with progressive 8K-to-32K-to-192K context extension and SFT for interleaved thinking followed by agentic RL; a January 2026 developer account specifically identifies CISPO-based multi-turn tool-use RL in the M2 generation. Original-checkpoint data volumes for each post-training stage, environment counts and total training expense remain undisclosed. Although the config advertises MTP, the original downloadable tensors contain no exported MTP modules. Launch-reported SWE-bench Verified is 69.4% using an OpenHands/R2E-Gym scaffold, 128K evaluation context, 100 steps and no test-time scaling; Terminal-Bench is 46.3%, averaged over eight runs with the original benchmark's Claude Code scaffold, not Terminal-Bench 2.0. Multi-SWE-Bench is 36.2% and SWE-bench Multilingual 56.5%, each averaged over eight Claude Code runs with a 300-step ceiling. These are tool-assisted scaffold results, not standalone coding scores or best-of-eight success rates. MiniMax's launch claim of 8% of Claude Sonnet 4.5's price and nearly twice its speed concerns its hosted API, not a universal local hardware or compute guarantee. This card covers only the original official checkpoint; community AWQ/GGUF conversions and later M2.1/M2.5/M2.7 releases are excluded.",
      "paper_url": "https://arxiv.org/abs/2605.26494",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M2",
        "https://modelscope.cn/models/MiniMax/MiniMax-M2",
        "https://github.com/MiniMax-AI/MiniMax-M2",
        "https://www.minimax.io/news/minimax-m2",
        "https://www.minimax.io/news/why-did-m2-end-up-as-a-full-attention-model",
        "https://www.minimax.io/news/post-training-experience-and-insights-for-agent-models",
        "https://github.com/MiniMax-AI/MiniMax-M2/blob/e27aac415872aa258b38315f4d51a751dbd31237/LICENSE",
        "https://github.com/MiniMax-AI/MiniMax-M2/blob/main/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m2/",
      "record_updated_at": "2026-10-11T08:37:45.385221+00:00"
    },
    {
      "slug": "ming-flash-omni-preview",
      "source_file": "2025-10-27-ming-flash-omni-preview.json",
      "release_date": "2025-10-27",
      "date": "October 27, 2025",
      "name": "Ming-flash-omni Preview",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "100B total, 6B active per token in the current model card; the technical report describes 100B / 6.1B, while the launch blog describes 103B / 9B",
      "context_window": "32,768 tokens",
      "license": "Hugging Face metadata identifies MIT, but the repository has no root LICENSE file and the complete incorporated-component licensing was not independently resolved.",
      "key_features": "A sparse any-to-any multimodal preview built on the Ling-Flash-2.0 MoE architecture. It accepts text, images, video and audio and generates text, speech and images. The release adds streaming video conversation, contextual and dialect-aware speech recognition, voice cloning, high-fidelity text rendering in generated images, stronger scene and identity consistency during image editing, and generative segmentation-as-editing. The publisher reports improvements across all twelve ContextASR benchmarks and recognition support for fifteen Chinese dialects. Its shipped language-model configuration has 32 transformer layers, 256 routed experts with eight selected per token plus one shared expert, grouped-query attention, and 32,768 maximum position embeddings. Forty-two root BF16 weight shards and the associated connector, speech and diffusion assets are publicly downloadable from Hugging Face and ModelScope. The later Ming-flash-omni-2.0 release is a separate successor and is not grouped into this card.",
      "paper_url": "https://arxiv.org/abs/2510.24821",
      "links": [
        "https://huggingface.co/inclusionAI/Ming-flash-omni-Preview",
        "https://www.modelscope.cn/models/inclusionAI/Ming-flash-omni-Preview",
        "https://github.com/inclusionAI/Ming",
        "https://www.inclusion-ai.org/blog/ming-flash-omni-preview"
      ],
      "page_url": "https://nlp.com.ai/models/ming-flash-omni-preview/",
      "record_updated_at": "2026-10-11T08:37:45.382699+00:00"
    },
    {
      "slug": "qwen3-vl-2b-32b-instruct-thinking",
      "source_file": "2025-10-21-qwen3-vl.json",
      "release_date": "2025-10-21",
      "date": "October 21, 2025",
      "name": "Qwen3-VL-2B / 32B (Instruct / Thinking)",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "Dense 2B / 32B; full BF16 checkpoints: 2,127,532,032 / 33,357,390,064 parameters, including vision encoder and four vision-language mergers",
      "context_window": "262,144 tokens native; optional extension to 1M with YaRN/configuration overrides",
      "license": "Apache 2.0 (explicit model-family weight grant and per-checkpoint hub declarations; standalone LICENSE absent from these HF packages)",
      "key_features": "Four separately released BF16 checkpoints: 2B-Instruct, 2B-Thinking, 32B-Instruct and 32B-Thinking. October 21 is the publisher-stated release date, retained provisionally: contemporaneous 2B local-installation evidence corroborates availability by that date, but exact first public access for all four is unresolved. Related FP8 packages are excluded from this same-day grouping pending historical timing and packaging review; the original 32B-Instruct-FP8 upload duplicated Thinking weights and was replaced on October 22. Dense Qwen3 text decoders accept text, images and sampled video frames and generate text; no native audio. Architecture combines interleaved MRoPE, textual video timestamps and DeepStack visual features injected into the first three decoder layers. The 2B decoder has 28 layers, width 2048, 16 query/8 KV heads and tied embeddings; its 24-layer, width-1024 vision encoder uses DeepStack indexes 5/11/17. The 32B decoder has 64 layers, width 5120, 64 query/8 KV heads and untied embeddings; its 27-layer, width-1152 vision encoder uses indexes 8/16/24. Native 256K context is shared across text, visual tokens and generation; advertised 1M requires positional-extension overrides, not the stock configuration. Publisher-described applications include OCR/document parsing, visual grounding, screenshot-to-code and GUI/tool-assisted agents; actual screen capture and actions require an external runtime. The retrospective technical report describes merger-only alignment (67B tokens), two all-parameter pretraining stages (~1T tokens each), then 100B-token 256K adaptation; exact per-checkpoint budgets are not separately disclosed. Post-training describes staged 32K-to-256K SFT with separate standard and long-CoT formats, strong-to-weak student distillation using teacher responses then on-policy KL logit alignment, and reasoning/general RL with SAPO. Small-model distillation is explicitly supported for 2B; the report does not establish an identical student recipe or exact optimization budget for every 32B checkpoint or the 235B teacher. Reported AIME-25 / LiveCodeBench v6 scores are 22.2/20.3 for 2B-Instruct, 39.0/29.3 for 2B-Thinking, 66.2/43.8 for 32B-Instruct and 83.7/65.6 for 32B-Thinking. Thinking uses an 81,920-token output budget on these reasoning evaluations, not an 81,920-token context limit. Reasoning is not universally better: 32B-Thinking scores 60.5 on GPT-4.1-judged Arena-Hard v2 versus Instruct 64.7; its LiveCodeBench advantage over text-only Qwen3-32B coexists with lower CFEval/OJBench scores. Reported hours-long video needle retrieval at 1M tests 235B-A22B-Instruct, not these sizes; no equivalent 2B/32B guarantee or universal minimal-VRAM claim is established.",
      "paper_url": "https://arxiv.org/abs/2511.21631",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-VL-2B-Thinking",
        "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-2B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-2B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-32B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-32B-Thinking",
        "https://github.com/QwenLM/Qwen3-VL"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-vl-2b-32b-instruct-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.380245+00:00"
    },
    {
      "slug": "qwen3-vl-4b-8b-instruct-and-thinking",
      "source_file": "2025-10-15-qwen3-vl.json",
      "release_date": "2025-10-15",
      "date": "October 15, 2025",
      "name": "Qwen3-VL-4B / 8B (Instruct and Thinking)",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "4B / 8B dense language backbones; 4,437,815,808 / 8,767,123,696 total checkpoint parameters including vision encoder and mergers",
      "context_window": "262,144 tokens native (shared text/image/video context); up to 1M advertised with explicit YaRN extension, not the default",
      "license": "Apache 2.0 (checkpoint declarations and explicit family weight grant; standalone LICENSE absent from the four weight repositories)",
      "key_features": "Four separate BF16 checkpoints pair dense Qwen3-derived 4B and 8B decoders with Instruct or reasoning-focused Thinking post-training. Both decoders have 36 layers and grouped-query attention with 32 query heads and 8 KV heads; the 4B ties its input/output embeddings, whereas the 8B does not. The 4B uses a 24-layer, width-1024 SigLIP2-Large-derived vision encoder with DeepStack features from layers 5/11/17; the 8B uses a 27-layer, width-1152 SigLIP2-SO-400M-derived encoder with features from layers 8/16/24. A main vision-language merger plus three DeepStack mergers integrate multilevel visual features; interleaved MRoPE and explicit textual video timestamps support spatial-temporal modeling. Inputs include text, images and sampled video frames, with generated text, document parsing/OCR, 2D/3D grounding, image-to-code and GUI-agent use cases; native audio input/output is not supported, and GUI execution requires an external agent runtime. The retrospective November 26 technical report describes merger-only alignment followed by full-parameter multimodal training at 8,192, 32,768 and 262,144 tokens; post-training uses standard or long-CoT SFT, strong-to-weak distillation for lightweight students using teacher responses and on-policy student/teacher KL alignment, then reasoning/general RL. Its family-level token budgets are not separately audited 4B/8B training totals. Reported size-specific Thinking scores are AIME-25 74.5/80.3 and LiveCodeBench v6 51.3/58.6 for 4B/8B, with an 81,920-token output allowance on these evaluations; corresponding Instruct scores are 46.6/45.9 and 37.9/39.3. Thinking is not uniformly better: 8B BFCL-v3 is 63.0 versus Instruct 66.3. The publisher advertises expanded multilingual OCR and optional 1M context, but its 32-of-39-language OCR threshold and 1M long-video needle-retrieval headline do not establish equivalent small-model accuracy or lossless full-resolution video comprehension. Official FP8 versions are public related packages, excluded from this BF16 release grouping because their exact public-download date is unresolved. The October 15 date is provisional: the official download-linked release statement was committed October 14 UTC, and contemporary October 15 local 4B execution corroborates availability by then, not the exact first unlock. The October 4 30B-A3B and October 21 2B/32B releases are separate events.",
      "paper_url": "https://arxiv.org/abs/2511.21631",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-VL-4B-Thinking",
        "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-VL-8B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-4B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-4B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-8B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-8B-Thinking",
        "https://github.com/QwenLM/Qwen3-VL",
        "https://github.com/QwenLM/Qwen3-VL/blob/a0d6248355509926085904f65c56a13a08f20bee/README.md",
        "https://arxiv.org/html/2511.21631v1"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-vl-4b-8b-instruct-and-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.378422+00:00"
    },
    {
      "slug": "ring-1t",
      "source_file": "2025-10-13-ring-1t.json",
      "release_date": "2025-10-13",
      "date": "October 13, 2025",
      "name": "Ring-1T",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "1T total, approximately 50B active per token (MoE)",
      "context_window": "65,536 tokens by default; up to 128K with YaRN extension",
      "license": "MIT (publisher-declared; no standalone weight LICENSE file in the checked packages)",
      "key_features": "Full post-trained trillion-scale text reasoning release, offered in BF16 and official FP8 packages; distinct from September's Ring-1T-preview and its Ling-1T-base foundation. Uses the Ling 2.0 architecture: 80 Transformer layers, GQA with 64 query heads and 8 KV heads, four initial dense layers, then top-8 routing among 256 experts plus one shared expert. The released packages have no additional MTP prediction layers. Training comprises long-CoT SFT on mathematics, STEM, code and other reasoning trajectories, followed by reasoning RLVR and general RLHF. The dedicated report describes IcePop, a GRPO-derived policy objective that calibrates token-level training/inference probability discrepancies and masks excessive-deviation updates, plus C3PO++ budget-controlled rollout scheduling and the ASystem training infrastructure; general RL uses GRPO. Publisher-reported results are AIME 2025 93.40%, HMMT 2025 86.72%, LiveCodeBench-v6 (2408-2505) 78.30%, CodeForces rating 2088 and ARC-AGI-1 55.94%, with the principal reasoning benchmarks evaluated at YaRN-extended 128K. In the separate AWorld multi-agent IMO 2025 evaluation, it solved Problems 1, 3, 4 and 5 on the first attempt, described as silver-medal-level; a nearly complete Problem 2 proof required a third attempt, and Problem 6 remained incorrect. This is scaffold-assisted natural-language reasoning, not a standalone checkpoint medal or a tool-assisted symbolic-solver result. The chat template supports tool-call serialization, but the report identifies advanced agentic/tool-use skills as under-optimized. Published limitations include identity confusion, language mixing, repetitive generation and substantial long-context GQA inference cost. The 128K extension requires inference configuration rather than being enabled in the default checkpoint.",
      "paper_url": "https://arxiv.org/abs/2510.18855",
      "links": [
        "https://huggingface.co/inclusionAI/Ring-1T",
        "https://modelscope.cn/models/inclusionAI/Ring-1T",
        "https://huggingface.co/inclusionAI/Ring-1T-FP8",
        "https://modelscope.cn/models/inclusionAI/Ring-1T-FP8",
        "https://github.com/inclusionAI/Ring-V2",
        "https://x.com/AntLingAGI/status/1977767599657345027"
      ],
      "page_url": "https://nlp.com.ai/models/ring-1t/",
      "record_updated_at": "2026-10-11T08:37:45.376372+00:00"
    },
    {
      "slug": "ling-1t",
      "source_file": "2025-10-09-ling-1t.json",
      "release_date": "2025-10-09",
      "date": "October 9, 2025",
      "name": "Ling-1T",
      "developer": "Ant Group / inclusionAI (China)",
      "parameters": "1T total MoE (999,705,309,184 learned checkpoint parameters); approximately 50B active per token (technical report: 51B)",
      "context_window": "32,768 tokens by default; up to 131,072 with explicitly configured YaRN",
      "license": "MIT (publisher hub declarations; no checkpoint LICENSE file or explicit weight-specific grant located)",
      "key_features": "Flagship text-only, non-thinking instruction model in Ling 2.0, distinct from the Ring reasoning series; its supplied chat template fixes detailed thinking off and supports tool-call formatting. The BF16 checkpoint has 80 layers, width 8,192, standard GQA with 64 query and 8 KV heads, QK normalization, half-head RoPE, and SwiGLU experts. The first four layers are dense; each subsequent MoE layer selects eight of 256 routed experts plus one shared expert, using sigmoid scores and auxiliary-loss-free bias balancing. This is not MLA or a hybrid linear-attention architecture. Pre-training used over 20T tokens, with reasoning-data share increasing from 32% to 46%; mid-training added long-context and curated chain-of-thought data, while the WSM scheduler merged checkpoints instead of conventional learning-rate decay. The Ling technical report describes decoupled supervised fine-tuning, Evo-CoT reinforcement learning with sentence-level LPO, and Group Arena Reward/RubriX preference alignment. Syntax–Function–Aesthetics rewards use rendered front-end screenshots as training feedback, not native image input. FP8 mixed-precision training and a one-layer MTP auxiliary objective are reported, but this released instruction checkpoint has no packaged MTP layers. Publisher evaluations report AIME 2025 70.42%, LiveCodeBench 61.68%, ArtifactsBench 59.31, and BFCL-V3 69.64%; comparisons include DeepSeek-V3.1-Terminus in non-thinking mode and Kimi-K2-Instruct-0905, and do not show universal superiority. The publisher acknowledges costly long-context GQA, limited multi-turn/agentic ability, and occasional instruction or identity deviations. October 9 is retained provisionally as an official open-source launch/available-by date, not a proven first public weight-download date. The separately downloadable ModelScope Ling-1T-FP8 package has an unresolved public-release date and is not silently grouped as same-day; the reported Ling-1T-base and Ring-preview ancestor inclusionAI/Ling-1T-base-2.0 were not verified as public checkpoints.",
      "paper_url": "https://arxiv.org/abs/2510.22115",
      "links": [
        "https://huggingface.co/inclusionAI/Ling-1T",
        "https://modelscope.cn/models/inclusionAI/Ling-1T",
        "https://modelscope.cn/models/inclusionAI/Ling-1T-FP8",
        "https://github.com/inclusionAI/Ling-V2",
        "https://www.antgroup.com/en/news-media/press-releases/1759982400000"
      ],
      "page_url": "https://nlp.com.ai/models/ling-1t/",
      "record_updated_at": "2026-10-11T08:37:45.374257+00:00"
    },
    {
      "slug": "glm-46",
      "source_file": "2025-09-30-glm-4-6.json",
      "release_date": "2025-09-30",
      "date": "September 30, 2025",
      "name": "GLM-4.6",
      "developer": "Zhipu AI (Z.ai, China)",
      "parameters": "355B total / 32B active (MoE, nominal); 352.798B main-model learned parameters, 356.786B including packaged MTP",
      "context_window": "200K tokens (202,752 configured positions)",
      "license": "MIT (weights); Apache-2.0 (repository code)",
      "key_features": "Text-only GLM-4.5 successor for coding, reasoning and tool-using agents. This candidate covers the original BF16 post-trained package; the related official FP8 package is linked for reference, but same-day public weight access remains unresolved and requires separate release consideration rather than an unqualified September grouping. No GLM-4.6 Base or Air release verified. The released architecture has 92 Transformer layers, hidden size 5,120, grouped-query attention with 96 query heads and 8 KV heads (not MLA), three initial dense layers, and MoE layers selecting 8 of 160 routed experts plus one shared expert. The BF16 package includes an auxiliary multi-token-prediction layer; the original FP8 package omitted MTP, which the publisher added on October 16. Hybrid thinking/non-thinking are runtime modes of the same checkpoint: thinking is enabled by default, while enable_thinking=false adds /nothink and an empty thinking prefix. Tool-integrated reasoning supports externally executed Python and browser/search tools; tool definitions and results must be supplied by the application. Context increases from GLM-4.5's 128K to 200K; the hosted 128K maximum output is distinct from checkpoint context. Z.ai reports launch scores of 68.0% SWE-bench Verified, 40.5% Terminal-Bench, 75.9% weighted tau2-Bench, 93.9% AIME 2025, 81.0% GPQA and 82.8% LiveCodeBench v6, evaluated at 128K context. With tools, it reports 98.6% AIME, 82.9% GPQA, 84.5% LiveCodeBench and 30.4% HLE versus 17.2% without tools. These are publisher evaluations, with agent scaffolds, tool access and incompletely disclosed attempt budgets, not universal standalone performance. The launch blog reports a 48.6% win rate against Claude Sonnet 4 in interactive human-assisted CC-Bench coding tasks and about 15% fewer tokens than GLM-4.5, while acknowledging weaker coding than Sonnet 4.5; later service documentation gives different efficiency claims. Improvements also target front-end code, writing, role-play and translation. GLM-4.6-specific continuation-training token counts, data mixture and SFT/RL stages are not disclosed in the checked primary sources; the cited GLM-4.5 report is background, not a dedicated GLM-4.6 training report.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/zai-org/GLM-4.6",
        "https://huggingface.co/zai-org/GLM-4.6-FP8",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.6",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.6-FP8",
        "https://z.ai/blog/glm-4.6",
        "https://github.com/zai-org/GLM-4.5",
        "https://docs.z.ai/guides/llm/glm-4.6"
      ],
      "page_url": "https://nlp.com.ai/models/glm-46/",
      "record_updated_at": "2026-10-11T08:37:45.371925+00:00"
    },
    {
      "slug": "ring-1t-preview",
      "source_file": "2025-09-29-ring-1t-preview.json",
      "release_date": "2025-09-29",
      "date": "September 29, 2025",
      "name": "Ring-1T-preview",
      "developer": "Ant Group / inclusionAI (China)",
      "parameters": "Approximately 1T total (MoE); approximately 50B active per token on the Ling-1T backbone",
      "context_window": "32,768 tokens configured; preview-specific validated maximum unreported",
      "license": "MIT (official weight-repository declaration; linked code license is also MIT)",
      "key_features": "Early open-weight trillion-scale reasoning checkpoint in the Ling 2.0 family, distinct from the subsequently trained October Ring-1T release. The preview's official metadata names inclusionAI/Ling-1T-base-2.0 as its ancestor, although that exact base repository could not be verified. Its text-only causal decoder has 80 layers, width 8,192, grouped-query attention with 64 query and 8 key/value heads, four initial dense layers, and MoE layers with 256 routed experts, eight selected per token and one shared expert. The released BF16 checkpoint contains 155 safetensors shards; its inference configuration has no extra MTP prediction layers, despite MTP being part of the broader Ling 2.0 design. The publisher says its foundation was pretrained on 20T tokens and the preview underwent reasoning-oriented reinforcement learning with verifiable rewards in ASystem, using IcePop; AReaL is the open-source RL component. The later Ring-1T report explains IcePop as GRPO-style policy optimization with token-level training/inference probability calibration and masking of excessive discrepancies, not merely synthetic trajectory generation, but the preview's exact SFT stages, reward mixture, RL budget and training cost are not disclosed. Publisher-reported early natural-language results include AIME 2025 92.6, HMMT 2025 84.5, LiveCodeBench v6 78.3 and ARC-AGI-1 50.8; the preview announcement does not fully specify sampling, aggregation or attempt budgets, and these are not independent downloaded-checkpoint evaluations. Its IMO 2025 Problem 3 success in one attempt was obtained inside the AWorld multi-agent framework, with partial solutions on Problems 1, 2, 4 and 5, not an unscaffolded single-model gold-medal result. The publisher explicitly notes language mixing, repetitive reasoning and identity misperception. The preview config sets max_position_embeddings=32768 and rope_scaling=null; later Ring-1T's 64K-to-128K YaRN support and newer 1M-context successors are not established preview limits. September 29 is the official UTC launch/download-link date, corresponding to September 30 in China; exact first public weight access remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/inclusionAI/Ring-1T-preview",
        "https://modelscope.cn/models/inclusionAI/Ring-1T-preview",
        "https://x.com/AntLingAGI/status/1972711364876697612",
        "https://x.com/AntLingAGI/status/1972711553578479701",
        "https://github.com/inclusionAI/Ling-V2",
        "https://github.com/inclusionAI/AWorld",
        "https://arxiv.org/abs/2510.18855"
      ],
      "page_url": "https://nlp.com.ai/models/ring-1t-preview/",
      "record_updated_at": "2026-10-11T08:37:45.368841+00:00"
    },
    {
      "slug": "deepseek-v32-exp",
      "source_file": "2025-09-29-deepseek-v3-2-exp.json",
      "release_date": "2025-09-29",
      "date": "September 29, 2025",
      "name": "DeepSeek-V3.2-Exp",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B backbone / approximately 37B active (MoE; additional DSA indexers; approximately 685B packaged with MTP)",
      "context_window": "128K supported (131,072 tokens; config positional ceiling 163,840)",
      "license": "MIT",
      "key_features": "Experimental text-generation successor to DeepSeek-V3.1-Terminus, introducing DeepSeek Sparse Attention (DSA) while retaining the 61-layer MLA/MoE backbone: 256 routed experts, eight selected per token, and one shared expert. A lightning indexer with 64 heads of dimension 128 selects up to 2,048 prior key-value tokens per query; core attention changes from quadratic to O(Lk), but the cheaper indexer remains quadratic. Original downloadable weights use mixed FP8/BF16 storage with UE8M0 quantization-scale configuration and include an auxiliary multi-token-prediction module. Continued pretraining starts from the 128K Terminus base checkpoint: a 2.1B-token dense-attention warm-up trains only the indexer with KL alignment, followed by 943.7B tokens of sparse training of all parameters, using separate indexer KL and main-model language-model objectives. Post-training matches the Terminus pipeline: RL-trained domain specialists supply thinking/non-thinking distillation data, followed by a single mixed GRPO stage combining reasoning, agent tasks and human alignment. The post-trained checkpoint supports thinking and non-thinking through its chat template, plus tool-call serialization for externally executed tools; these are modes of the same weights, not separate variants. Reported quality is broadly comparable to Terminus, not uniformly improved: AIME 2025 pass@1 89.3 versus 88.4, GPQA-Diamond 79.9 versus 80.7, Humanity’s Last Exam 19.8 versus 21.7, and LiveCodeBench (August 2024–May 2025) 74.1 versus 74.9. Tool-assisted SWE Verified is 67.8 versus 68.4 and BrowseComp 40.1 versus 38.5; these depend on agent scaffolds and are not unaided chat scores. The report attributes some reasoning regressions to fewer generated reasoning tokens. Long-context inference savings are measured on deployed H800 clusters assuming US$2 per GPU-hour, not inferred from API price reductions; real-world sparse-attention limitations remained under validation. This card covers the post-trained experimental checkpoint; the separately published V3.2-Exp-Base is related, but exact same-day public-download grouping remains unconfirmed. Later December V3.2/Speciale are excluded. The official November 17 inference-demo update fixes indexer RoPE layout that could degrade earlier local runs.",
      "paper_url": "https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp",
        "https://github.com/deepseek-ai/DeepSeek-V3.2-Exp",
        "https://api-docs.deepseek.com/news/news250929",
        "https://api-docs.deepseek.com/zh-cn/news/news250929"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v32-exp/",
      "record_updated_at": "2026-10-11T08:37:45.365946+00:00"
    },
    {
      "slug": "qwen3-vl-235b-a22b",
      "source_file": "2025-09-23-qwen3-vl-235b-a22b.json",
      "release_date": "2025-09-23",
      "date": "September 23, 2025",
      "name": "Qwen3-VL-235B-A22B",
      "developer": "Alibaba / Qwen Team (China)",
      "parameters": "235B total / 22B active per text token (MoE decoder); 235,670,022,896 learned parameters including vision encoder and four mergers; original Instruct and Thinking checkpoints",
      "context_window": "256K (262,144 tokens), native shared text/image/video context; optional YaRN extension to approximately 1M tokens",
      "license": "Apache 2.0 (checkpoint model-card declarations; standalone LICENSE files absent from the checked weight repositories)",
      "key_features": "Original flagship Qwen3-VL release with separate non-thinking Instruct and reasoning-enhanced Thinking checkpoints, accepting text, images and video frames and generating text; no native audio input or output. The Qwen3 decoder has 94 layers, grouped-query attention with 64 query heads and 4 KV heads, and 128 routed experts with 8 selected per token, without shared experts. Its modified SigLIP2-SO-400M-initialized vision encoder has 27 layers, width 1,152 and 16-pixel patches; a main merger compresses 2×2 visual patches, while three additional DeepStack mergers inject features from ViT layers 8/16/24 into the first three decoder layers without adding extra visual tokens. Interleaved-MRoPE balances temporal and spatial rotary frequencies, and explicit textual timestamps improve video event localization. Supports multilingual OCR, structured document parsing, multi-image dialogue, normalized 2D grounding, 3D spatial reasoning and screenshot/sketch-to-HTML/CSS/JavaScript or Draw.io generation. The launch advertises usable OCR in 32 languages, with over 70% accuracy in 32 of a 39-language test beyond Chinese and English. Visual-agent and tool-assisted perception capabilities require an external runtime to execute tools and GUI actions. The later Qwen3-VL report describes merger-only alignment followed by joint text/multimodal pretraining at 8K, 32K and 256K sequence lengths, then standard-answer or chain-of-thought SFT at 32K and 256K and reasoning/general reinforcement learning; strong-to-weak distillation is specifically described for lightweight students, not established as a flagship training stage. Exact launch-checkpoint stage budgets and variant-specific RL allocation remain incompletely disclosed. Publisher-reported flagship Thinking/Instruct scores include MMMU-Pro 69.3/68.1, MathVista mini 85.8/84.9 and Video-MME without subtitles 79.0/79.2; comparison models use different reasoning budgets, input settings and sometimes externally sourced results, so these are not universal superiority claims. The report's Instruct video needle-retrieval test achieves 100% at native 256K (about 30 minutes) and 99.5% with YaRN-expanded 1M (about two hours), using 1-FPS sampling and resolution adjusted to a visual-token budget—not guaranteed lossless full-resolution video recall. Official FP8 packages and smaller October models are later releases and are excluded from this September grouping. September 23 is the publisher-declared weight-release date; exact first public unlock remains unresolved, with contemporaneous local checkpoint handling corroborated by September 24 and successful execution by September 25.",
      "paper_url": "https://arxiv.org/abs/2511.21631",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-235B-A22B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-VL-235B-A22B-Thinking",
        "https://github.com/QwenLM/Qwen3-VL",
        "https://qwen.ai/blog?id=99f0335c4ad9ff6153e517418d48535ab6d8afef"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-vl-235b-a22b/",
      "record_updated_at": "2026-10-11T08:37:45.363047+00:00"
    },
    {
      "slug": "qwen3-omni-30b-a3b-instruct-thinking",
      "source_file": "2025-09-22-qwen3-omni.json",
      "release_date": "2025-09-22",
      "date": "September 22, 2025",
      "name": "Qwen3-Omni-30B-A3B (Instruct / Thinking)",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "Thinker: nominal 30B total / 3B active; full Instruct checkpoint: 35,259,818,545 stored parameters; Thinking: 31,719,205,488; Instruct adds a nominal 3B / 0.3B-active Talker and speech-generation modules",
      "context_window": "65,536 tokens configured for Thinker; Instruct Talker also 65,536 (code predictor: 32,768)",
      "license": "Apache 2.0 (explicit checkpoint/model-family grant in the technical report; standalone LICENSE files absent from the three HF repositories)",
      "key_features": "Native multimodal Thinker–Talker MoE family accepting text, images, audio and video, including temporally aligned video audio. Instruct produces text and streaming speech; the separate Thinking checkpoint contains only the Thinker and produces chain-of-thought text, not speech. Select the checkpoint explicitly rather than assuming a shared thinking-mode toggle. The Thinker has 48 layers, 128 routed experts with eight selected per token, GQA and interleaved multimodal rotary positions; the Talker uses 128 routed experts with six selected plus a shared expert. AuT audio and SigLIP2-initialized vision encoders supplement the language decoder, so 30B/3B does not describe the whole speech pipeline. Speech generation uses 16 discrete codec codebooks, residual-code prediction and a lightweight causal Code2Wav ConvNet instead of block-wise diffusion, with 12.5-Hz codec frames. The report describes 119 text languages, 19 speech-input languages and 10 speech-output languages. Original Instruct exposes three voices: Chelsie, Ethan and Aiden. Publisher-reported theoretical cold-start first-packet latency is 234 ms for audio-only and 547 ms for audiovisual input at concurrency one with optimized vLLM; this is not a guaranteed network-inclusive response time. Pretraining aligns encoders/adapters to a frozen Qwen3-initialized LLM, then jointly trains on roughly 2T multimodal tokens and extends training sequences from 8,192 to 32,768. AuT was separately trained on 20M hours of supervised/pseudo-labeled audio. Thinker post-training combines SFT, off-/on-policy teacher distillation and GSPO with rule/model rewards; Talker uses multimodal speech training, high-quality continued pretraining, multilingual DPO and speaker fine-tuning. Training sequence lengths and reported 30–40-minute audio handling are not checkpoint context-window guarantees. In publisher evaluations, AIME25 is 65.0 for Instruct and 73.7 for Thinking, while BFCL-v3 is 64.4 and 63.2 respectively; evaluation uses greedy Instruct decoding, checkpoint sampling settings for Thinking, no system prompt and two video frames per second. The headline 32/36 open-source and 22/36 overall audio/audiovisual SOTA claim covers the evaluated suite including unreleased in-house Flash variants, not a guarantee for every downloadable checkpoint. The report identifies weaker long-video performance due to context and positional-extrapolation limits. Related same-launch Captioner is an Instruct-derived, single-turn audio-only/text-output fine-tune accepting no user text prompts, with clips of at most 30 seconds recommended; it is referenced but excluded from this text-prompt-capable grouping. Hosted Flash, experimental Base comparisons and later 2512 checkpoints are not included. September 22 is verified availability-by from the explicit public release/download materials and contemporary open-weight reporting; the exact first public unlock remains unresolved.",
      "paper_url": "https://arxiv.org/abs/2509.17765",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-Omni-30B-A3B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-Omni-30B-A3B-Thinking",
        "https://github.com/QwenLM/Qwen3-Omni",
        "https://qwen.ai/blog?id=qwen3-omni",
        "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Captioner",
        "https://modelscope.cn/models/Qwen/Qwen3-Omni-30B-A3B-Captioner"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-omni-30b-a3b-instruct-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.359903+00:00"
    },
    {
      "slug": "deepseek-v31-terminus",
      "source_file": "2025-09-22-deepseek-v3-1-terminus.json",
      "release_date": "2025-09-22",
      "date": "September 22, 2025",
      "name": "DeepSeek-V3.1-Terminus",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B total / 37B active (MoE; approximately 685B including packaged MTP module)",
      "context_window": "128K supported (163,840-token configuration ceiling)",
      "license": "MIT",
      "key_features": "Single updated, text-only V3.1 post-trained checkpoint built on the existing DeepSeek-V3.1-Base, not a new Terminus Base release. Retains the V3 architecture: 61 transformer layers, Multi-head Latent Attention (MLA), three initial dense layers followed by MoE layers with 256 routed experts, eight selected per token and one shared expert; one auxiliary multi-token prediction (MTP) layer is packaged separately. It does not use the DeepSeek Sparse Attention introduced later in V3.2-Exp. The same weights support thinking and non-thinking responses via the chat-template thinking flag: generation prefixes end in <｜Assistant｜><think> or <｜Assistant｜></think>, respectively. General tool calls use the non-thinking template; the updated search-agent trajectory supplies a specialized reasoning/search format and revised tool set. DeepSeek reports reduced Chinese-English mixing and abnormal characters, plus improved Code Agent and Search Agent performance. Launch reasoning-without-tools results include GPQA-Diamond 80.7 and text-only Humanity's Last Exam 21.7, versus V3.1's 80.1 and 15.9. Agentic tool-assisted results improve SWE-bench Verified from 66.0 to 68.4, multilingual SWE-bench from 54.5 to 57.8, Terminal-bench from 31.3 to 36.7, BrowseComp from 30.0 to 38.5 and SimpleQA from 93.4 to 96.8; the SimpleQA score is a search-agent result, not unaided factual accuracy. Improvements are not uniform: BrowseComp-zh falls from 49.2 to 45.0, Codeforces from 2091 to 2046 and Aider-Polyglot from 76.3 to 76.1. Search results depend on the internal search scaffold, commercial search API and webpage filtering; SWE results use an internal code-agent framework, and Terminal-bench's Terminus 1 harness is unrelated to the checkpoint's Terminus name. Launch sampling, retry and token-budget details are not fully disclosed. Inherits V3.1-Base's long-context continuation (630B tokens at 32K and 209B at 128K), not a newly announced Terminus pretraining budget; the update's complete post-training data and optimization recipe remain undisclosed. Official mixed FP8/BF16 weights use E4M3 blocks and UE8M0-compatible power-of-two scales stored as FP32 tensors, but the original release explicitly warns that self_attn.o_proj scales do not conform to UE8M0; compatible inference kernels and FP32 routing correction biases matter. Public downloads are explicitly offered in the September 22 launch; exact first public-unlock timing remains unverified.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Terminus",
        "https://api-docs.deepseek.com/news/news250922",
        "https://api-docs.deepseek.com/zh-cn/news/news250922",
        "https://github.com/deepseek-ai/DeepSeek-V3"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v31-terminus/",
      "record_updated_at": "2026-10-11T08:37:45.357286+00:00"
    },
    {
      "slug": "qwen3-next-80b-a3b-instruct-thinking",
      "source_file": "2025-09-11-qwen3-next-80b-a3b.json",
      "release_date": "2025-09-11",
      "date": "September 11, 2025",
      "name": "Qwen3-Next-80B-A3B (Instruct / Thinking)",
      "developer": "Alibaba (Qwen Team) (China)",
      "parameters": "80B total / approximately 3B active per token (sparse MoE; 79.674B main-model learned parameters, 81.325B packaged including MTP)",
      "context_window": "262,144 tokens native (256K); up to 1,010,000 with explicit YaRN scaling",
      "license": "Apache 2.0",
      "key_features": "Original text-only Qwen3-Next release with two distinct post-trained checkpoints: Instruct is non-thinking only, while Thinking always uses reasoning and its chat template pre-fills <think>; these are not an enable_thinking toggle on one checkpoint. The 48-layer hybrid alternates three Gated DeltaNet layers with one gated full-attention layer (36 linear / 12 full), each followed by sparse MoE with 512 routed experts, 10 selected per token and one shared expert. Hidden width is 2,048; full attention uses 16 query / 2 KV heads of dimension 256 with partial RoPE, while DeltaNet uses 16 QK / 32 value heads of dimension 128. Qwen reports pretraining on a uniformly sampled 15T-token subset of its 36T-token Qwen3 corpus, followed by post-training; GSPO is explicitly credited for stabilizing Thinking RL, but the complete checkpoint-specific SFT/RL recipe and trained context schedule are not disclosed in the checked launch sources. Zero-centered RMSNorm, weight decay on normalization weights and normalized router initialization improve training stability. Multi-token prediction is trained to improve model quality and supplies an auxiliary speculative-decoding module; acceleration requires supported inference implementations and is not generally available in Transformers. Qwen claims base-model pretraining used 9.3% of Qwen3-32B compute cost (GPU hours) and that prefill/decode throughput exceeds Qwen3-32B by 10x beyond 32K context; the published comparison does not disclose hardware, precision or batch size and is not an end-to-end cost or universal latency guarantee. Publisher benchmarks report Instruct/Thinking GPQA scores of 72.9/77.2, AIME25 of 69.5/87.8 and LiveCodeBench v6 (February-May 2025) of 56.6/68.7; exact score-specific sampling and tool protocols are not fully specified, so these are not verified no-tool results. Instruct approaches the much larger Qwen3-235B-A22B-Instruct-2507 on selected tasks and improves long-context RULER, but does not match it uniformly; Thinking can require substantially longer generation. Native context is a shared input/output budget; the advertised approximately 1M extension requires YaRN factor 4 and an inference length override, with possible short-context degradation. Original downloads are BF16, with the full approximately 80B weights needed despite sparse activation; official FP8 packages appeared later and are not same-day launch variants. The evaluated original Base is not verified as a public launch checkpoint, and later Coder-Next and other derivatives are excluded. September 11 is provisional: official GitHub release news and contemporary local-use evidence support availability by that date, but the official download-linked blog displays September 10; exact first public access and the source calendar/timezone discrepancy remain unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking",
        "https://modelscope.cn/models/Qwen/Qwen3-Next-80B-A3B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-Next-80B-A3B-Thinking",
        "https://qwen.ai/blog?id=qwen3-next"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-next-80b-a3b-instruct-thinking/",
      "record_updated_at": "2026-10-11T08:37:45.354543+00:00"
    },
    {
      "slug": "kimi-k2-instruct-0905",
      "source_file": "2025-09-05-kimi-k2-instruct-0905.json",
      "release_date": "2025-09-05",
      "date": "September 5, 2025",
      "name": "Kimi K2-Instruct-0905",
      "developer": "Moonshot AI (China)",
      "parameters": "1T total / 32B active per token (MoE; nominal)",
      "context_window": "256K tokens (262,144)",
      "license": "Modified MIT; retain copyright/permission notices, and prominently display \"Kimi K2\" in the UI of commercial products/services exceeding 100M monthly active users OR US$20M monthly revenue",
      "key_features": "September update to the text-only, non-long-thinking Kimi K2 Instruct model, separate from the July Base/Instruct release and November K2 Thinking. Retains a 61-layer decoder with one dense layer, Multi-head Latent Attention (MLA), 384 routed experts, eight selected per token and one shared expert; no auxiliary MTP layers are configured. Doubles the supported context from July's 128K to 256K. Moonshot emphasizes stronger agentic coding, more attractive and practical frontend/UI code, tool calling and integration with coding-agent scaffolds; graphical demos are generated code, not evidence of vision input. The September release table reports SWE-bench Verified 69.2 ± 0.63%, Multilingual 55.9 ± 0.72%, Multi-SWE-Bench 33.5 ± 0.28%, Terminal-Bench 44.5 ± 2.03% and SWE-Dev 66.6 ± 0.72%, as mean ± standard deviation over five independent full-test-set runs. Most use Moonshot's modified SWE-agent-derived harness with restricted Bash/Edit context and task-specific prompts; Terminal-Bench uses Terminus-2. These are scaffold-dependent results, not best-of-five selection or OpenHands scores; September maximum-turn and per-task retry budgets are not specified. July's 65.8% SWE-bench single-attempt and 71.6% multi-attempt results must not be substituted for this update. Inherits K2's 15.5T-token MuonClip-pretrained foundation and its agentic post-training background, but the amount, data mix and exact continued-training/post-training recipe for 0905 are undisclosed. Official weights are block-FP8 with BF16 tensors and FP32 scales, not a separately verified BF16 edition. Native tool parsing and the matching tokenizer/chat template are required; tool IDs follow functions.func_name:idx with a globally incrementing counter, and Moonshot documents multi-turn failures when IDs or runtime versions are wrong. Despite sparse 32B activation, serving must accommodate the trillion-parameter weight set: the supplied 256K H200 deployment guide specifies 16 GPUs. Remaining benchmark gaps, integration fragility and deployment cost qualify the coding improvements; hosted turbo speed and format-enforcement guarantees are service properties, not local-weight guarantees.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905",
        "https://modelscope.cn/models/moonshotai/Kimi-K2-Instruct-0905",
        "https://platform.kimi.com/blog/posts/kimi-k2-0905",
        "https://github.com/MoonshotAI/Kimi-K2",
        "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/blob/main/LICENSE",
        "https://arxiv.org/abs/2507.20534v1"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k2-instruct-0905/",
      "record_updated_at": "2026-10-11T08:37:45.351849+00:00"
    },
    {
      "slug": "wayfarer-2-12b",
      "source_file": "2025-09-04-wayfarer-2-12b.json",
      "release_date": "2025-09-04",
      "date": "September 4, 2025",
      "name": "Wayfarer-2-12B",
      "developer": "LatitudeGames (USA)",
      "parameters": "12B (12,247,802,880 parameters; dense)",
      "context_window": "131,072 tokens (configured; finetune supported maximum unverified)",
      "license": "Apache 2.0",
      "key_features": "English text-adventure and roleplay finetune of Mistral-Nemo-Base-2407, developed by Latitude with Gryphe Padar; not a Llama model or a confirmed continuation of the original Wayfarer weights. Dense text-only Mistral causal transformer with 40 layers, grouped-query attention (32 query heads/8 KV heads), 5,120 hidden width and 131,074-token checkpoint vocabulary. Latitude describes the sequel as slower-paced, with longer, more detailed responses, frequent failure and potential death for all characters to counter positivity bias. SFT combines the Wayfarer 2 adventure dataset, sentiment-balanced roleplay transcripts and a small instruction core. The card describes synthetic adventure simulations with separate narrator and player models, five user archetypes per scenario and runs of up to 8K tokens or until the protagonist dies; this is not the checkpoint context limit. Uses ChatML and AI Dungeon-style \"> You...\" action inputs. Trained for second-person present-tense narration; other viewpoints can be suboptimal, and the publisher acknowledges weaker instruction following outside AI Dungeon-oriented formatting. Official product guidance also warns of repetition and difficulty maintaining third person. Recommended checkpoint sampling baseline is temperature 0.8, repetition penalty 1.05 and min_p 0.025, not the different hosted-service presets. Public BF16 safetensors and publisher-endorsed GGUF packages are available. Configured context is 128K, but the finetune training window and validated long-context maximum are undisclosed; hosted AI Dungeon limits are separate. No release-specific benchmarks, DPO/reward-training recipe or technical paper were verified.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Wayfarer-2-12B",
        "https://huggingface.co/LatitudeGames/Wayfarer-2-12B-GGUF",
        "https://www.reddit.com/r/LocalLLaMA/comments/1n8kk48/",
        "https://help.aidungeon.com/ai-model-differences"
      ],
      "page_url": "https://nlp.com.ai/models/wayfarer-2-12b/",
      "record_updated_at": "2026-10-11T08:37:45.349175+00:00"
    },
    {
      "slug": "nova-70b-llama-33",
      "source_file": "2025-09-04-nova-70b-llama-3-3.json",
      "release_date": "2025-09-04",
      "date": "September 4, 2025",
      "name": "Nova-70B-Llama-3.3",
      "developer": "Latitude (USA)",
      "parameters": "70B (dense; 70,553,706,496 parameters)",
      "context_window": "131,072 tokens configured; validated finetune maximum unknown",
      "license": "Llama 3.3 Community License Agreement + Acceptable Use Policy (custom, not Apache 2.0)",
      "key_features": "English, text-only narrative and roleplay finetune trained directly on Meta Llama 3.3 70B Instruct, not on Wayfarer Large; dense decoder-only Transformer with 80 layers, 8,192 hidden width, grouped-query attention with 64 query heads and 8 key/value heads, and a 128,256-token vocabulary. Developed with Gryphe Padar for character consistency, relationships, emotional depth, complex narratives and nuanced instruction-following. Training combines text adventures of the kind used for Wayfarer, long emotional narratives, detailed worldbuilding and general roleplay, carefully balanced and rewritten to remove common AI cliches, plus a small single-turn instruction dataset. Latitude reports that additional DPO experiments worsened performance when combined with Llama 3.3's existing DPO training; the released checkpoint's complete optimization recipe, dataset sizes and training sequence lengths are not disclosed. Trained exclusively on second-person present-tense narrative data, although the publisher says the larger instruct backbone can handle other perspectives and complex instructions. Official service guidance identifies possible interview-style dialogue, flowery prose and stalled scenes; no quantitative Nova-specific benchmark is published in the release card. Uses original Llama 3 role headers and end-of-turn tokens, not ChatML. Model-card sampling baseline is temperature 0.8, repetition penalty 1.05 and min_p 0.05; bundled generation defaults instead use temperature 0.6 and top_p 0.9. Public BF16 weights and publisher-endorsed GGUF quantizations support local inference without an AI Dungeon subscription. Configured 128K capacity is not a published validation of Nova's maximum effective context, and hosted subscription windows are separate service limits. September 4 is the publisher's public open-weight release declaration; exact first public download timing remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Nova-70B-Llama-3.3",
        "https://huggingface.co/LatitudeGames/Nova-70B-Llama-3.3-GGUF",
        "https://www.reddit.com/r/LocalLLaMA/comments/1n8kk48/new_ai_dungeon_models_wayfarer_2_12b_nova_70b/",
        "https://help.aidungeon.com/ai-model-differences",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/USE_POLICY.md"
      ],
      "page_url": "https://nlp.com.ai/models/nova-70b-llama-33/",
      "record_updated_at": "2026-10-11T08:37:45.346549+00:00"
    },
    {
      "slug": "command-a-translate",
      "source_file": "2025-08-28-command-a-translate.json",
      "release_date": "2025-08-28",
      "date": "August 28, 2025",
      "name": "Command A Translate",
      "developer": "Cohere / Cohere Labs (Canada)",
      "parameters": "111B (dense; 111,057,580,032 parameters)",
      "context_window": "8K input + 8K output (16K advertised total)",
      "license": "CC-BY-NC-4.0 with Acceptable Use Addendum + Cohere Labs Acceptable Use Policy; noncommercial open weights, commercial use requires a separate agreement",
      "key_features": "Translation-specialized, text-only autoregressive model fine-tuned from Command A, covering 23 languages. Uses a dense 64-layer Cohere2 decoder with grouped-query attention (96 query/8 KV heads), tied embeddings, and repeating three 4,096-token sliding-window/RoPE layers followed by one global-attention layer without positional embeddings. The dedicated report says the 111B translation adaptation uses DPO only, not an additional SFT stage: 126,000 difficulty- and quality-filtered documents averaging 951 tokens, with original human translations preferred over Command A-generated alternatives and language balancing including non-English pairs. Natural-language translation requests use the standard chat template; no specialized target-language tags are required. Cohere reports leading WMT24++ translation quality: launch xCometXL 83.8 (84.4 with multi-step Deep Translation); the later report gives 83.9 (84.9 with Deep Translation), using English-to-target paragraph-level evaluation rather than testing every language pair. These are automatic metrics, not a verified 65% human win rate. Deep Translation is a separate iterative inference workflow, and the two-model CommandA-WMT submission is not this checkpoint. Reported long-document quality falls from 56.9 xCometXL with paragraph-by-paragraph translation to 51.9 in one-shot long-context translation, below Gemini 2.5 Pro’s 56.2 in that setting. Prompt-injection translation failure is 0.3% under the report’s question-mark heuristic, not a general safety guarantee. Translation-focused performance does not establish general-chat, coding or agent benchmark quality; current API docs list tool use but launch-specific validation is not established. Cohere advertises private deployment and one A100/H100 with 4-bit quantization at less than 0.5 xCometXL quality loss; the original open checkpoint is BF16. Public research downloads require accepting the Hugging Face noncommercial terms. August 28 is verified availability-by from the official launch offer; exact first public unlock remains unresolved.",
      "paper_url": "https://aclanthology.org/2025.wmt-1.55/",
      "links": [
        "https://huggingface.co/CohereLabs/command-a-translate-08-2025",
        "https://cohere.com/blog/command-a-translate",
        "https://docs.cohere.com/docs/command-a-translate",
        "https://cohere.com/c4ai-cc-by-nc-license",
        "https://docs.cohere.com/docs/cohere-labs-acceptable-use-policy"
      ],
      "page_url": "https://nlp.com.ai/models/command-a-translate/",
      "record_updated_at": "2026-10-11T08:37:45.343899+00:00"
    },
    {
      "slug": "hermes-4",
      "source_file": "2025-08-26-hermes-4.json",
      "release_date": "2025-08-26",
      "date": "August 26, 2025",
      "name": "Hermes 4",
      "developer": "Nous Research (USA)",
      "parameters": "70B / 405B (dense; exact learned totals: 70,553,706,496 / 405,853,388,800)",
      "context_window": "131,072 tokens configured (128K); validated maximum for the finetunes not independently established",
      "license": "Llama 3.1 Community License Agreement plus its Acceptable Use Policy for both sizes, including FP8 derivatives; not Apache 2.0 or unrestricted MIT. Redistribution requires the agreement, notice and Built with Llama attribution; distributed derived model naming has a Llama-prefix requirement. Additional permission applies above 700M MAU under the agreement's release-date condition. HF metadata says llama3, but the repos do not package a LICENSE file; parent obligations remain applicable.",
      "key_features": "Original text-only hybrid-reasoning release: Hermes-4-70B and Hermes-4-405B, with official BF16 and FP8 packaging. Both are Llama 3.1 derivatives according to the launch model cards and technical report; the 70B is not documented as Llama 3.3. August 26 is a provisional public-weight availability-by date, not a proven first unlock; August 25 paper submission and earlier uploads do not establish public access. The Qwen3-derived 14B was explicitly released for download September 3 and is excluded from this same-day grouping; no official Hermes-4-32B was verified. Dense decoder-only transformers use grouped-query attention (70B: 80 layers, width 8192, 64 query/8 KV heads; 405B: 126 layers, width 16384, 128 query/8 KV heads). A single checkpoint supports direct answers or explicit <think>...</think> reasoning: thinking=True selects a reasoning system prompt, thinking defaults to false, and keep_cots=True preserves prior reasoning; custom system messages must include the reasoning instruction themselves. These Llama variants use Llama-3 role-header formatting, not ChatML; tool schemas and JSON calls use <tools>, <tool_call> and <tool_response> tags, with actual tool execution supplied by external software. Training combines synthetic DataForge instructions, retained Hermes 3 data and Atropos rejection-sampled trajectories checked by approximately 1,000 task verifiers. The report describes about 5M samples (3.5M reasoning plus 1.6M non-reasoning), 19B dataset tokens and 56B training tokens per size; model cards advertise about 60B tokens versus Hermes 3's 1.2B, a roughly 50-fold token expansion, not 50% more data. Documented training is assistant-only cross-entropy SFT with packed samples on 192 B200 GPUs at a 16,384-token training window; no DPO/PPO/GRPO policy-training stage is established for these two checkpoints. Published evaluation uses 40,960-token reasoning/code contexts and 32,768 for other tasks, not a validation of the full configured 128K. In reasoning mode, 405B/70B score 81.9/73.5 on AIME 2024, 78.1/67.4 on AIME 2025, 70.5/66.1 on GPQA Diamond and 61.3/50.5 on LiveCodeBench v6 (454 problems dated August 2024–May 2025). These estimate pass@1 using 64, 64, 8 and 16 samples respectively, not best-of-N success. The report finds strong reasoning, creative writing, persona steering and schema adherence, but substantial task/mode tradeoffs: 405B non-reasoning AIME 2025 is 10.6, and loose IFEval is 81.5 reasoning versus 84.9 non-reasoning (70B: 78.7 versus 82.3). SimpleQA accuracy is only 25.8/17.9 in reasoning mode, so factual reliability is not guaranteed; these are accuracy scores, not measured hallucination rates. RefusalBench favors reduced refusals except in three inverted safety categories and is not a general safety certification. No verified Hermes 4 BFCL score or controlled OpenAI o3 comparison was found in the original report; a blanket frontier-parity claim is unsupported.",
      "paper_url": "https://arxiv.org/abs/2508.18255v1",
      "links": [
        "https://huggingface.co/NousResearch/Hermes-4-70B",
        "https://huggingface.co/NousResearch/Hermes-4-405B",
        "https://huggingface.co/NousResearch/Hermes-4-70B-FP8",
        "https://huggingface.co/NousResearch/Hermes-4-405B-FP8",
        "https://huggingface.co/lmstudio-community/Hermes-4-70B-GGUF",
        "https://huggingface.co/lmstudio-community/Hermes-4-405B-GGUF",
        "https://x.com/NousResearch/status/1960416954457710982",
        "https://nousresearch.com/releases",
        "https://github.com/NousResearch/torchtitan/tree/856a0ecabeb8a882c150641f73f8c1c235720622",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/USE_POLICY.md"
      ],
      "page_url": "https://nlp.com.ai/models/hermes-4/",
      "record_updated_at": "2026-10-11T08:37:45.341292+00:00"
    },
    {
      "slug": "grok-2",
      "source_file": "2025-08-23-grok-2.json",
      "release_date": "2025-08-23",
      "date": "August 23, 2025",
      "name": "Grok-2",
      "developer": "xAI (USA)",
      "parameters": "269.515B total / approximately 114.897B active (MoE)",
      "context_window": "131,072 tokens (official configured maximum; validated supported limit undisclosed)",
      "license": "Custom Grok 2 Community License Agreement + xAI AUP; not Apache 2.0. Initial August 23 terms restricted commercial use to licensees and affiliates with annual revenue below US$1M; an August 24 UTC revision removed that threshold in favor of AUP compliance. Restrictions on training other foundation/general-purpose AI models and distribution attribution (including Powered by xAI) remain; current November 4, 2025 xAI Community License is a later revision.",
      "key_features": "Public weights of xAI’s post-trained Grok-2 used in 2024, released as xai-org/grok-2 rather than a separate Instruct alias. Text-only causal decoder: the released config and all weight tensors contain no vision/audio encoder, and hosted Grok-2 vision, image generation and real-time X retrieval are not included. The 64-layer, width-8,192 Transformer uses grouped-query attention with 64 query heads and 8 KV heads of dimension 128, eight routed experts with two selected per token, and a shared gated MLP alongside the routed MoE; this is not a dense 270B model. Audited tensors contain 269,515,497,472 learned parameters; approximately 114.897B active uses conventional counting that includes both full embedding/output tables. The official config sets 131,072 positions, original_max_position_embeddings=8,192 and scaling_factor=16, but publishes no checkpoint-specific long-context validation; these fields do not establish an 8K training limit or justify a community YaRN override. Weights are BF16 across 39 safetensors files, approximately 539 GB, with expert tensors presharded for TP=8; FP8 in the official SGLang serving command is runtime quantization, not the downloaded checkpoint dtype. Official usage requires SGLang >=0.5.1, eight GPUs each with over 40 GB memory, and its Human/Assistant chat template. xAI’s August 2024 hosted evaluation reported MMLU 87.5%, GPQA 56.0% and HumanEval pass@1 88.4%, but these are not independently validated scores for the August 2025 weight export. Full pretraining data/token budget and checkpoint-specific SFT/RL recipe are undisclosed, and local deployment has substantial storage/memory and specialized-runtime requirements. Only Grok-2 is included: Grok-2 mini weights were not verified, and Grok-1 is a separate release. August 23 is a provisional weight-release date supported by the publisher’s live-release statement and contemporary public repository activity; exact first public download/unlock remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/xai-org/grok-2",
        "https://huggingface.co/xai-org/grok-2/blob/main/config.json",
        "https://huggingface.co/xai-org/grok-2/blob/0533f67c37c21df05df948777975f8551fa62d48/LICENSE",
        "https://huggingface.co/xai-org/grok-2/blob/d60cbe267db8bb43be676bc80e200c64268ea8ec/LICENSE",
        "https://x.com/elonmusk/status/1959379349322313920",
        "https://x.ai/news/grok-2"
      ],
      "page_url": "https://nlp.com.ai/models/grok-2/",
      "record_updated_at": "2026-10-11T08:37:45.338784+00:00"
    },
    {
      "slug": "deepseek-v31",
      "source_file": "2025-08-21-deepseek-v3-1.json",
      "release_date": "2025-08-21",
      "date": "August 21, 2025",
      "name": "DeepSeek-V3.1",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B total / 37B activated per token (MoE; approximately 685B packaged including auxiliary MTP tensors)",
      "context_window": "128K supported (163,840-position configuration ceiling)",
      "license": "MIT",
      "key_features": "Original text-only post-trained DeepSeek-V3.1 checkpoint, not the later V3.1-Terminus or V3.2 releases. Its hybrid thinking/non-thinking modes use the same weights with different chat-template prefixes, not a hybrid attention architecture or separate Think checkpoint. It retains DeepSeek-V3's 61-layer Multi-head Latent Attention decoder and MoE structure: the first three layers use dense feed-forward networks; subsequent layers have 256 routed experts, eight selected per token, and one shared expert. V3.1-Base, the training ancestor, continues the original V3 base with long-context extension: 630B tokens at 32K and 209B at 128K, described as approximately 840B additional tokens in the official release. The post-trained model improves tool use and agent tasks; DeepSeek also reports chain-of-thought compression giving comparable average task quality to R1-0528 with 20–50% fewer output tokens. A complete V3.1-specific post-training/RL recipe is not published in the checked sources. Both original checkpoints advertise 128K context; the configured 163,840 positions do not establish a supported 160K context. The published FP8 weights use E4M3 with UE8M0-compatible power-of-two scales and dynamic activation quantization; correct scale handling and the revised tokenizer/tool-call parsers matter for deployment. Official recommendations require router correction biases in FP32 and UE8M0 scales for weights and activations; existing generic V3 FP8/runtime support is not a blanket compatibility guarantee. Thinking-mode publisher scores include GPQA Diamond 80.1%, AIME 2024 93.1%, AIME 2025 88.4%, LiveCodeBench (August 2024–May 2025) 74.8%, and text-only Humanity's Last Exam 15.9%; reasoning comparisons are not uniformly better than R1-0528, whose GPQA and HLE scores are 81.0% and 17.7%. Non-thinking agent scores are SWE-bench Verified 66.0%, SWE-bench Multilingual 54.5%, and Terminal-bench 31.3% with the Terminus 1 framework. Thinking search-agent results include BrowseComp 30.0%, BrowseComp_zh 49.2%, and HLE with Python plus search 29.8%; these use external tools and an internal commercial-search/webpage-filter framework with 128K context, while the R1 comparator uses a predefined workflow. SWE results use DeepSeek's internal code-agent scaffold, not unaided single-response coding. General tool calling is documented in non-thinking mode, with specialized thinking-mode search trajectories. August 21 is verified availability-by for the post-trained weights, not an exact first-unlock timestamp. The original Base weights were independently converted and run by August 20 UTC, so Base is an ancestry reference here rather than an asserted same-day grouped variant.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.1",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1",
        "https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Base",
        "https://www.deepseek.com/news/deepseek-v3-1/",
        "https://www.deepseek.com/en/news/deepseek-v3-1/",
        "https://github.com/deepseek-ai/DeepSeek-V3",
        "https://github.com/deepseek-ai/DeepGEMM"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v31/",
      "record_updated_at": "2026-10-11T08:37:45.336279+00:00"
    },
    {
      "slug": "command-a-reasoning",
      "source_file": "2025-08-21-command-a-reasoning.json",
      "release_date": "2025-08-21",
      "date": "August 21, 2025",
      "name": "Command A Reasoning",
      "developer": "Cohere / Cohere Labs (Canada)",
      "parameters": "111B dense (111,057,580,032 learned parameters)",
      "context_window": "256K supported (500,000 configured positional ceiling); 32K maximum output",
      "license": "CC BY-NC 4.0 with Cohere Labs Acceptable Use Addendum/AUP; commercial use requires separate permission",
      "key_features": "Original text-only, autoregressive Cohere2 research checkpoint command-a-reasoning-08-2025, derived from March Command A; not the July Vision checkpoint or a separate model for each reasoning mode. The 64-layer dense decoder uses width 12,288, GQA with 96 query heads and 8 KV heads, head dimension 128, and a tied 256,000-token embedding. Its hybrid attention repeats three 4,096-token sliding-window RoPE layers followed by a global-attention layer without positional embeddings. The launch post advertises 128K context for low-footprint single-GPU deployments and 256K with two or more GPUs; the current config's 500,000-position ceiling is not a validated 500K context claim. Cohere describes post-pretraining supervised fine-tuning and preference training for helpfulness and safety; a Reasoning-specific RL algorithm, reward design and training-data recipe are not disclosed in the checked primary sources. Trained for conversational tool use, JSON-schema tools, retrieval-augmented generation and agents, with optional grounded citations. Supports reasoning on/off within one checkpoint: the Hugging Face chat template uses reasoning=True by default or reasoning=False for direct responses, with <|START_THINKING|>/<|END_THINKING|> and <|START_RESPONSE|>/<|END_RESPONSE|> delimiters. The hosted API separately exposes an enforced thinking.token_budget that switches to the final response at the limit; the published local template does not itself implement a hard thinking-token budget. Cohere recommends leaving at least 1K output tokens for the answer, e.g. a 31K thinking budget within the 32K output limit; tool schemas, observations and conversation history also consume context. Cohere claims coverage of English plus 22 other languages, not equal measured quality in every language. In launch evaluations at the highest available reasoning setting, Cohere reports 70.3% BFCL-v3 using function calling, 59.8% tau-bench averaged over airline/retail and 10 pass^1 runs, and 46.5% multilingual tau-bench averaged over six languages including English, both domains and 10 runs, ahead of R1-0528, gpt-oss-120b and Magistral Medium in that protocol. These are task-specific enterprise-agent results, not blanket leadership on academic reasoning or coding. Its DeepResearchBench and North claims evaluate complete retrieval/multi-agent systems, not the isolated checkpoint. Reasoning increases latency and compute; 111B BF16 weights alone occupy about 222 GB, so advertised single-A100/H100 deployment should not be read as unquantized BF16 residency. Current production guidance instead recommends four H100 GPUs. June 2024 knowledge cutoff and imperfect agent scores remain relevant limitations; no model-specific factuality or GPQA/AIME/SWE-bench result with verified launch protocol was found.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/CohereLabs/command-a-reasoning-08-2025",
        "https://cohere.com/blog/command-a-reasoning",
        "https://docs.cohere.com/docs/command-a-reasoning",
        "https://docs.cohere.com/docs/reasoning",
        "https://cohere.com/cohere-labs-cc-by-nc-license"
      ],
      "page_url": "https://nlp.com.ai/models/command-a-reasoning/",
      "record_updated_at": "2026-10-11T08:37:45.334674+00:00"
    },
    {
      "slug": "seed-oss-36b",
      "source_file": "2025-08-20-seed-oss-36b.json",
      "release_date": "2025-08-20",
      "date": "August 20, 2025",
      "name": "Seed-OSS-36B",
      "developer": "ByteDance Seed Team (China)",
      "parameters": "36B dense (36,151,104,512 parameters each; Base, Base-woSyn and Instruct)",
      "context_window": "512K (524,288 tokens)",
      "license": "Apache 2.0",
      "key_features": "Text-only causal Transformer family with 64 layers, hidden width 5,120, 80 query heads and 8 key/value heads of dimension 128, RoPE, RMSNorm and SwiGLU; input embeddings and output head are untied. The original release comprises Seed-OSS-36B-Base, Seed-OSS-36B-Base-woSyn and Seed-OSS-36B-Instruct, not three parameter sizes. ByteDance reports roughly 12T pretraining tokens and native training up to 512K context. Base includes synthetic instruction data during pretraining; Base-woSyn omits that synthetic instruction component as a research ablation, not a guarantee that every training datum is human-authored. Instruct is post-trained from the synthetic-augmented Base. The model card identifies SFT and RLHF/PPO safety training, but a complete model-specific training recipe and staged sequence lengths were not verified. Instruct controls reasoning length through thinking_budget and <seed:cot_budget_reflect> self-reflection tokens inside <seed:think> blocks. Recommended budgets are 512-token multiples through 16K; 0 requests a direct answer, while the default -1 leaves reasoning without an explicit thinking budget, still bounded by generation and context limits. This is prompt/template-based control, not a hard token-count limiter; the original zero-budget template still starts a thinking/reflection block. The developer reports Instruct AIME24 91.7%, AIME25 84.7%, GPQA-Diamond 71.4%, HLE 10.1%, LiveCodeBench v6 (February-May 2025) 67.4%, and RULER at 128K 94.6%. Its coding-agent results depend on scaffolding: SWE-bench Verified 56% with OpenHands versus 47% with AgentLess 4*10. Main benchmark generation uses temperature 1.1/top_p 0.95, with different TAU settings; tool policy, thinking budgets and sample aggregation are not fully specified for all results. These are developer-reported comparisons, not universal superiority or proof of 512K retrieval quality. Base benefits from synthetic instruction data on MATH (81.7% versus woSyn 61.3%) but trails woSyn on GPQA-Diamond (31.7% versus 35.2%). Instruct remains weak on SimpleQA (9.7%) and ARC-AGI v2 (1.45%); its 85.8% IFEval is below several listed peers. Official limitations include hallucination and primarily English evaluation despite international-use positioning. August 20 is the provisional publisher-declared download release date; the corporate blog is dated August 21, contemporary local-use evidence is from August 21 UTC, and the exact first public unlock remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base",
        "https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base-woSyn",
        "https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct",
        "https://github.com/ByteDance-Seed/seed-oss",
        "https://seed.bytedance.com/en/blog/seed-oss-open-source-models-release"
      ],
      "page_url": "https://nlp.com.ai/models/seed-oss-36b/",
      "record_updated_at": "2026-10-11T08:37:45.332345+00:00"
    },
    {
      "slug": "gemma-3-270m",
      "source_file": "2025-08-14-gemma-3-270m.json",
      "release_date": "2025-08-14",
      "date": "August 14, 2025",
      "name": "Gemma 3 (270M)",
      "developer": "Google (USA)",
      "parameters": "270M nominal; 268,098,176 total dense parameters per PT/IT checkpoint (167,772,160 tied embedding parameters + 100,326,016 other parameters)",
      "context_window": "32K (32,768 tokens, shared input/output budget)",
      "license": "Custom Gemma Terms of Use + incorporated Gemma Prohibited Use Policy (not Apache 2.0 or an unrestricted permissive license)",
      "key_features": "Original August 2025 text-only Gemma 3 270M release, comprising pretrained and instruction-tuned checkpoints, separate from the March larger Gemma 3 models and Gemma 3n. Dense decoder-only Transformer with 18 layers, hidden width 640, intermediate width 2,048, four query heads sharing one key/value head (head dimension 256), and a repeated five-local/one-global attention pattern; local sliding windows span 512 tokens. A 262,144-token vocabulary uses tied input/output embeddings; there is no image encoder. The official 270M model card reports pretraining on 6 trillion tokens and an August 2024 data cutoff; the precise 270M teacher/distillation and instruction-tuning stages are not separately documented in the March family report. Designed primarily as a compact foundation for task-specific fine-tuning, including classification, entity extraction, query routing and structured-text processing, rather than complex general-purpose conversation. Published out-of-the-box IT results include IFEval 51.2 (0-shot), ARC-c 28.2 (0-shot), and BIG-Bench Hard 26.7 (few-shot); PT TriviaQA is 15.4 (5-shot). These are checkpoint-specific benchmark results, not task-fine-tuned production scores. Google cautions that open-ended or complex tasks and factual/common-sense reliability remain limitations; broad family multilingual claims do not establish equivalent proficiency for this tiny checkpoint. Same-day official PT/IT QAT packages are unquantized BF16 checkpoints intended for subsequent Q4_0 quantization, not ready-made INT4 files. Google reports that an INT4 deployment used 0.75% battery for 25 conversations in internal Pixel 9 Pro SoC tests and demonstrates offline browser use with Transformers.js; these are configuration-specific deployment claims. August 14 public download availability is corroborated by the official download invitation and contemporaneous local use; exact first public-access timing is unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/google/gemma-3-270m",
        "https://huggingface.co/google/gemma-3-270m-it",
        "https://huggingface.co/google/gemma-3-270m-qat-q4_0-unquantized",
        "https://huggingface.co/google/gemma-3-270m-it-qat-q4_0-unquantized",
        "https://developers.googleblog.com/en/introducing-gemma-3-270m/",
        "https://ai.google.dev/gemma/docs/core/model_card_3",
        "https://ai.google.dev/gemma/docs/releases",
        "https://ai.google.dev/gemma/terms",
        "https://ai.google.dev/gemma/prohibited_use_policy"
      ],
      "page_url": "https://nlp.com.ai/models/gemma-3-270m/",
      "record_updated_at": "2026-10-11T08:37:45.330702+00:00"
    },
    {
      "slug": "gpt-oss-gpt-oss-120b-gpt-oss-20b",
      "source_file": "2025-08-05-gpt-oss.json",
      "release_date": "2025-08-05",
      "date": "August 5, 2025",
      "name": "gpt-oss (gpt-oss-120b / gpt-oss-20b)",
      "developer": "OpenAI (USA)",
      "parameters": "gpt-oss-120b: 116.83B total / 5.13B active per token; gpt-oss-20b: 20.91B total / 3.61B active per token (MoE)",
      "context_window": "131,072 tokens (both variants)",
      "license": "Apache License 2.0; accompanying gpt-oss usage policy requires compliance with applicable law",
      "key_features": "Two original text-only, autoregressive reasoning/chat MoE checkpoints released together; the 120b and 20b names round actual totals of about 117B and 21B. The larger model has 36 layers and 128 experts per layer, the smaller 24 layers and 32 experts; both route each token to four SwiGLU experts, use a 2,880-wide residual stream, and alternate 128-token sliding-window and full attention with GQA (64 query heads, eight key/value heads) and YaRN-scaled 131,072-token context. Released expert matrices use MXFP4 (4.25 bits per parameter including scales), while attention and other non-expert components remain BF16; packed storage elements and quantization scales are not additional learned parameters. OpenAI reports that optimized MXFP4 inference fits 120b on one 80GB GPU and 20b on systems with 16GB memory, not that BF16 inference or a fully populated 128K context fits those budgets; KV cache, runtime and batch size affect memory, and CPU speed is not established by these fit claims. Trained on trillions of text tokens emphasizing STEM, code and general knowledge, with a June 2024 knowledge cutoff and harmful-content filtering; post-training uses chain-of-thought reinforcement learning similar to o3, tool-use training and deliberative safety alignment. Exact corpus composition, token budget and full post-training recipe are not disclosed in the launch report. Low/medium/high reasoning effort is selected in the system prompt, not by loading separate weights. Harmony formatting is required, separating analysis, commentary/tool recipients and final answers; supports developer-defined function calls and structured outputs. Browsing and Python require external tools and an execution harness, not capabilities that run autonomously inside the weights. In OpenAI's high-effort, quantized evaluations, AIME 2025 scores are 92.5%/91.7% without tools and 97.9%/98.7% with tools (120b/20b); HLE is 14.9%/10.9% without tools and 19.0%/17.3% with tools. SWE-bench Verified scores of 62.4%/60.7% are tool-assisted evaluation results, not standalone code-generation or demonstrated live-agent reliability. Strengths include efficient reasoning, mathematics and tool-oriented coding; factual reliability is weaker: without browsing, SimpleQA hallucination rates are 78.2%/91.4%, versus 75.0% for o4-mini in the same report. Longer reasoning increases latency and cost. OpenAI evaluated adversarially fine-tuned 120b variants for biological/chemical and cyber risk and reported that they did not reach its High-capability thresholds under the tested conditions; this is not a guarantee against downstream misuse or safety removal. Later safeguard models, community derivatives and purported unquantized conversions are outside this release grouping.",
      "paper_url": "https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf",
      "links": [
        "https://openai.com/index/introducing-gpt-oss/",
        "https://huggingface.co/openai/gpt-oss-120b",
        "https://huggingface.co/openai/gpt-oss-20b",
        "https://github.com/openai/gpt-oss",
        "https://github.com/openai/harmony"
      ],
      "page_url": "https://nlp.com.ai/models/gpt-oss-gpt-oss-120b-gpt-oss-20b/",
      "record_updated_at": "2026-10-11T08:37:45.328522+00:00"
    },
    {
      "slug": "command-a-vision",
      "source_file": "2025-07-31-command-a-vision.json",
      "release_date": "2025-07-31",
      "date": "July 31, 2025",
      "name": "Command A Vision",
      "developer": "Cohere and Cohere Labs (Canada)",
      "parameters": "112B total (111,867,525,360 checkpoint parameters; 111B text backbone)",
      "context_window": "128K supported; model card describes a 32K Hugging Face default",
      "license": "CC BY-NC 4.0 with Cohere acceptable-use addendum/AUP; non-commercial research weights (commercial use requires separate permission)",
      "key_features": "Single dense image-and-text-to-text research checkpoint, command-a-vision-07-2025, built on Command A with a SigLIP2-so400m-patch16-512 vision encoder and an MLP multimodal adapter; not a Pixtral encoder or a mixture-of-experts model. Cohere2Vision uses the Command A text architecture; the available mirror configuration specifies 64 decoder layers, grouped-query attention with 96 query/8 key-value heads, and repeating three 4,096-token sliding-attention layers followed by one full-attention layer. Images are mapped to aspect-ratio-aware 512x512 tiles: up to 12 tiles plus a thumbnail, pixel-shuffled to 256 visual tokens per tile and at most 3,328 tokens per image; the model card recommends images up to 2048x1536 (3 megapixels). Supports 128K context, not the text-only parent's 256K; the card's 32K default and the mirror's 500,000-position ceiling are not separate validated context claims. Optimized for document OCR, chart/diagram interpretation, in-image tables and scene understanding, with six explicitly supported languages: English, Portuguese, Italian, French, German and Spanish. Cohere's hosted API allows up to 20 images or 20MB per request and 8K output; these are service limits rather than independently established local-checkpoint image/output maxima. Inputs are text and images, outputs are text; official documentation says tool use is unsupported and it does not generate images. Training comprises frozen-encoder/frozen-language-model vision-language alignment, joint encoder/adapter/language-model supervised fine-tuning, multimodal expert merging, then regularization and RLHF including online Contrastive Policy Gradient; detailed data mixtures, training-token budgets and full hyperparameters are not published in the launch description. Cohere's launch evaluation reports ChartQA 90.9%, InfoVQA 82.9%, AI2D 94.0%, MMMU with CoT 65.3%, MathVista 73.5%, DocVQA 95.9%, TextVQA 84.8%, OCRBench 86.9% and RealWorldQA 73.6% (nine-task average 83.1%). Strong chart/document/OCR results do not imply uniformly best reasoning: GPT-4.1 and Llama 4 Maverick score higher on MMMU, and GPT-4.1 on RealWorldQA in the same table. Comparisons mix reported scores with best-effort internal evaluations, typically VLMEvalKit, without a complete matched image-resolution/prompting protocol. A Cohere author cautions that languages outside the six targeted languages may perform worse despite the multilingual text backbone. Public research weights require accepting an automatic Hugging Face non-commercial/contact-sharing gate; commercial API access has separate terms. July 31 is verified public-weight availability-by through the official dated release and local-loading instructions, not independently proven first public unlock timing.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/CohereLabs/command-a-vision-07-2025",
        "https://huggingface.co/blog/CohereLabs/introducing-command-a-vision-07-2025",
        "https://cohere.com/blog/command-a-vision",
        "https://docs.cohere.com/docs/command-a-vision",
        "https://docs.cohere.com/v1/changelog/2025-07-31-command-a-vision",
        "https://cohere.com/c4ai-cc-by-nc-license"
      ],
      "page_url": "https://nlp.com.ai/models/command-a-vision/",
      "record_updated_at": "2026-10-11T08:37:45.326901+00:00"
    },
    {
      "slug": "glm-45",
      "source_file": "2025-07-28-glm-4-5.json",
      "release_date": "2025-07-28",
      "date": "July 28, 2025",
      "name": "GLM-4.5",
      "developer": "Zhipu AI (Z.ai, China)",
      "parameters": "GLM-4.5: 355B total / 32B active; GLM-4.5-Air: 106B total / 12B active (MoE; publisher counts include MTP but exclude embeddings and output layer)",
      "context_window": "128K (131,072 tokens; GLM-4.5 and GLM-4.5-Air)",
      "license": "MIT (weights; explicit launch/model-card declarations; standalone checkpoint LICENSE files absent)",
      "key_features": "Text-only agentic, reasoning, and coding family, released as GLM-4.5 and the smaller GLM-4.5-Air, with BF16 Base and post-trained checkpoints plus official FP8 post-trained packages. The post-trained weights support thinking and non-thinking responses in the same checkpoint; thinking is enabled by default, and the launch chat template supports enable_thinking=false and /nothink. Hybrid reasoning refers to these response modes, not a linear-attention hybrid architecture or native visual input. Both use full grouped-query attention with partial RoPE, 96 query heads and 8 KV heads, sigmoid-gated MoE routing with 8 selected routed experts and 1 shared expert per MoE layer. GLM-4.5 has 92 main layers (3 dense, 89 MoE), 160 routed experts per MoE layer, and QK normalization; Air has 46 main layers (1 dense, 45 MoE), 128 routed experts, and no QK normalization. Each package also includes one auxiliary MoE multi-token-prediction layer for speculative decoding; stored checkpoint tensor totals should not be confused with the advertised embedding-excluded parameter counts. The dedicated report describes about 23T tokens for GLM-4.5: 15T general pretraining, 7T code/reasoning continuation, and subsequent repository-level code, synthetic reasoning, and long-context/agent mid-training, extending sequence length from 4K through 32K to 128K. Post-training uses cold-start SFT, reasoning/agent/general-chat expert RL, iterative self-distillation, unified expert-distillation SFT with mixed reasoning and direct-response data, and final general RL for instruction following, function calling, human/AI preferences, and pathological-output correction; the published sources do not separately establish Air's complete token budget. Z.ai reports SWE-bench Verified scores of 64.2%/57.6% for GLM-4.5/Air using OpenHands v0.34.0, a 100-iteration horizon and history truncation, not standalone code generation; its report does not specify a repeated-attempt aggregation for these scores. Terminal-Bench scores are 37.5%/30.0% using Terminus with function calling, and BFCL v3 Full scores are 77.8%/76.4%. Main-model reasoning results include AIME 2024 91.0% (average over 32 samples), GPQA 79.1% (average over 8 samples), and text-only HLE 14.4%; these averages are not best-of-N selection. Strong tool-use and software-engineering results do not imply uniformly superior factual accuracy: the report gives SimpleQA 26.4%/14.5%. Local serving requires substantial GPU memory, and full-context deployment needs more memory than short-context inference. Native image/video capabilities of the later GLM-4.5V and improvements from GLM-4.6 or later releases are not included.",
      "paper_url": "https://arxiv.org/abs/2508.06471",
      "links": [
        "https://z.ai/blog/glm-4.5",
        "https://github.com/zai-org/GLM-4.5",
        "https://huggingface.co/zai-org/GLM-4.5",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5",
        "https://huggingface.co/zai-org/GLM-4.5-Air",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5-Air",
        "https://huggingface.co/zai-org/GLM-4.5-FP8",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5-FP8",
        "https://huggingface.co/zai-org/GLM-4.5-Air-FP8",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5-Air-FP8",
        "https://huggingface.co/zai-org/GLM-4.5-Base",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5-Base",
        "https://huggingface.co/zai-org/GLM-4.5-Air-Base",
        "https://modelscope.cn/models/ZhipuAI/GLM-4.5-Air-Base"
      ],
      "page_url": "https://nlp.com.ai/models/glm-45/",
      "record_updated_at": "2026-10-11T08:37:45.324700+00:00"
    },
    {
      "slug": "qwen3-coder-480b-a35b-instruct",
      "source_file": "2025-07-22-qwen3-coder-480b-a35b-instruct.json",
      "release_date": "2025-07-22",
      "date": "July 22, 2025",
      "name": "Qwen3-Coder-480B-A35B-Instruct",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "480B total (480,154,875,392 learned parameters); 35B active per token; MoE with 160 routed experts, 8 selected, no shared expert",
      "context_window": "262,144 tokens native; up to 1M advertised with YaRN extrapolation (inference configuration required)",
      "license": "Apache 2.0",
      "key_features": "Original text-only, non-thinking agentic coding checkpoint, released in BF16 and an official blockwise FP8 package; not the later 30B-A3B or Qwen3-Coder-Next models. Its 62-layer decoder uses grouped-query attention with 96 query heads and 8 KV heads, QK normalization, and routed MoE feed-forward layers, not MLA or hybrid linear attention. Supports chat, code generation and fill-in-the-middle; the launch repository advertises 358 coding languages/formats. The original release describes pretraining on 7.5T tokens with 70% code, synthetic-data cleaning/rewriting using Qwen2.5-Coder, execution-driven large-scale Code RL, and long-horizon agentic RL with infrastructure supporting 20,000 independent environments in parallel. The precise SFT/RL algorithms and stage budgets were not disclosed in the launch blog. Native 256K context supports repository-scale work; advertised 1M support requires YaRN extrapolation and is not enabled by the released null rope_scaling configuration. Publisher launch results include SWE-bench Verified 69.6% with OpenHands and a 500-turn horizon, versus 67.0% at 100 turns; Claude Sonnet 4 scores 70.4% and 68.0% in those corresponding configurations. These are tool-assisted agent results, claimed without test-time scaling, not standalone code-generation pass@1 or best-of-eight scores. Other publisher results include Aider-Polyglot 61.8, Terminal-Bench 37.5, WebArena 49.9 and BFCL-v3 68.7; evaluation details beyond the launch table are incomplete, and superiority is task-dependent. Browser, terminal and coding-agent actions require external tools/scaffolds such as Qwen Code, Cline or configured Claude Code integrations; they are not built-in browser access or a vision modality. Uses a dedicated Qwen3-Coder function-call format/parser and tokenizer; unlike original hybrid Qwen3, it has no supported thinking toggle. Practical limitations include the full 480B-weight resident/offload burden despite 35B active compute, KV-cache cost at long context, and runtime/parser compatibility. FP8 uses 128-by-128 weight blocks with dynamic activations; quantization scales are storage overhead, not additional learned parameters. July 22 is a provisional launch/download-offer date: exact first public weight access and the July 22 UTC versus July 23 publisher-local timing remain unresolved for parent review.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct",
        "https://modelscope.cn/models/Qwen/Qwen3-Coder-480B-A35B-Instruct",
        "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8",
        "https://modelscope.cn/models/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8",
        "https://qwenlm.github.io/blog/qwen3-coder/",
        "https://github.com/QwenLM/Qwen3-Coder",
        "https://github.com/QwenLM/qwen-code"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-coder-480b-a35b-instruct/",
      "record_updated_at": "2026-10-11T08:37:45.322959+00:00"
    },
    {
      "slug": "hyperclova-x-seed-14b-think",
      "source_file": "2025-07-22-hyperclova-x-seed-14b-think.json",
      "release_date": "2025-07-22",
      "date": "July 22, 2025",
      "name": "HyperCLOVA X SEED 14B Think",
      "developer": "NAVER Cloud (South Korea)",
      "parameters": "14.748B (14,748,112,896; dense, untied embeddings)",
      "context_window": "32K (published supported context; 131,072 configured positional ceiling)",
      "license": "HyperCLOVA X SEED 14B Think Model License Agreement + Acceptable Use Policy (custom; commercial use subject to restrictions; separate NAVER permission required for >10M preceding-month MAU as of release or substantially similar/directly competing products/services; attribution and derivative-name obligations)",
      "key_features": "Original text-only Korean/English reasoning and instruction-following checkpoint, compressed from the HyperCLOVA X THINK flagship through structured pruning and knowledge distillation; not a Llama-weight derivative, sparse MoE, or vision model. Its Llama-style Transformer uses 38 layers, 6,144 hidden dimensions, grouped-query attention with 48 query/8 KV heads, Peri-Layer Normalization and maximal update parameterization (μP). The launch card describes SFT on reasoning tasks, RL with verifiable rewards, reasoning-length controllability, then joint RLHF/RLVR. The cited THINK-family report describes modified GRPO and correctness, format, language and length rewards; exact 14B-specific data quantities and all flagship hyperparameters are not independently disclosed. One checkpoint supports both thinking and direct-answer modes: the launch template defaults to model-selected reasoning, force_reasoning=True forces thinking, and skip_reasoning=True requests a direct answer; force_reasoning takes precedence. Natural-language reasoning-budget requests are approximate, not hard output limits. ChatML tool/function-call formatting supports external-tool integration, but tools must be executed by the application. NAVER's launch evaluations report a Korean non-thinking benchmark average of 0.7269 versus Qwen3-14B's 0.7079, but an English average of 0.6614 versus 0.6807. In thinking evaluations, KoBALT is 0.4500 versus 0.3840 and HAERAE 0.8537 versus 0.7410; MATH500 is 0.9380 versus 0.9680 and MBPP 0.8759 versus 0.9080, so advantages are task-specific. Sampling/aggregation details for these 14B tables are not fully specified. NAVER reports 68,049 A100-80GB-equivalent backbone GPU hours at assumed 50% MFU, 52.60×/91.38× below its Qwen2.5-14B/Qwen3-14B estimates; this is not independently audited end-to-end development cost. Only the original 14B Think release is covered; thinking modes are not separate Base/non-thinking variants. July 22 is supported by NAVER's dated release/download invitation and launch license, but the exact first public weight-unlock time remains unresolved.",
      "paper_url": "https://arxiv.org/abs/2506.22403",
      "links": [
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-14B",
        "https://www.navercorp.com/media/pressReleasesDetail?seq=33144",
        "https://clova.ai/en/tech-blog/a-new-seed-grows-in-the-ai-ecosystem-hyperclova-x-seed-think",
        "https://github.com/NAVER-Cloud-HyperCLOVA-X/hcx-vllm-plugin"
      ],
      "page_url": "https://nlp.com.ai/models/hyperclova-x-seed-14b-think/",
      "record_updated_at": "2026-10-11T08:37:45.320736+00:00"
    },
    {
      "slug": "qwen3-235b-a22b-instruct-2507",
      "source_file": "2025-07-21-qwen3-235b-a22b-instruct-2507.json",
      "release_date": "2025-07-21",
      "date": "July 21, 2025",
      "name": "Qwen3-235B-A22B-Instruct-2507",
      "developer": "Alibaba Qwen Team (China)",
      "parameters": "235B total / 22B active (MoE; 235,093,634,560 learned parameters)",
      "context_window": "262,144 tokens natively; later optional extension to 1,010,000 tokens",
      "license": "Apache 2.0",
      "key_features": "July 2025 update of the flagship Qwen3 non-thinking model: a text-only causal Transformer with 94 layers, grouped-query attention (64 query heads and 4 KV heads), and 128 routed experts with 8 selected per token and no shared experts. Unlike the April hybrid checkpoint, this release supports only non-thinking mode, does not emit <think></think> blocks, and needs no enable_thinking=False switch; it can nevertheless produce step-by-step answers and use tools. Qwen reports stronger instruction following, logical reasoning, mathematics, science, coding and tool use, improved multilingual long-tail knowledge, and better preference alignment for subjective and open-ended writing. Its launch card reports MMLU-Pro 83.0, GPQA 77.5, IFEval 88.7, BFCL-v3 70.9, AIME25 70.3, LiveCodeBench v6 (February-May 2025) 51.8, and Arena-Hard v2 79.2 (GPT-4.1-judged win rate). These are publisher evaluations, not uniformly documented pass@1 results or independent measurements; tool benchmarks depend on their evaluation setup. Improvements are not universal: Aider-Polyglot falls to 57.3 from 59.6 for the April non-thinking model, and TAU-Airline 44.0 trails Claude Opus 4 non-thinking at 59.6. Recommended sampling is temperature 0.7, top-p 0.8, top-k 20, with a 16,384-token output budget; higher presence penalties can reduce repetition but introduce language mixing. Native launch context is 256K; the 1,010,000-token configuration and specialized long-context attention instructions were added on August 8, not part of the July launch. Official BF16 and blockwise FP8 packages represent the same checkpoint, not distinct models. The July-specific training-token budget and detailed alignment recipe are undisclosed in the checked primary sources; the original Qwen3 report describes the April family, not this update. Separate July 25 Thinking-2507 weights, later smaller 2507 sizes, and Qwen3-Coder are excluded. July 21 is the publisher-declared release date; exact first public weight access remains unresolved.",
      "paper_url": "https://arxiv.org/abs/2505.09388",
      "links": [
        "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507",
        "https://modelscope.cn/models/Qwen/Qwen3-235B-A22B-Instruct-2507",
        "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8",
        "https://modelscope.cn/models/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8",
        "https://github.com/QwenLM/Qwen3",
        "https://qwenlm.github.io/blog/qwen3/"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-235b-a22b-instruct-2507/",
      "record_updated_at": "2026-10-11T08:37:45.318411+00:00"
    },
    {
      "slug": "openreasoning-nemotron",
      "source_file": "2025-07-16-openreasoning-nemotron.json",
      "release_date": "2025-07-16",
      "date": "July 16, 2025",
      "name": "OpenReasoning-Nemotron",
      "developer": "NVIDIA (USA)",
      "parameters": "1.5B / 7B / 14B / 32B (four dense models; checkpoint tensor counts approximately 1.544B / 7.616B / 14.770B / 32.764B)",
      "context_window": "64K declared at launch; 131,072 tokens configured in all four checkpoints, with supported maximum unresolved",
      "license": "CC-BY-4.0 for all four released models; Apache 2.0 applies to their Qwen2.5 reference models, not as the sole OpenReasoning weight license",
      "key_features": "Text-only decoder-only reasoning models based on Qwen2.5 at four sizes, specialized for competitive mathematics, coding and scientific problem solving rather than general-purpose chat. NVIDIA reports supervised fine-tuning distillation only, without reinforcement learning, using approximately 5 million reasoning solutions generated by DeepSeek-R1-0528 671B from math, code and science questions. Model cards identify OpenMathReasoning, OpenCodeReasoning/OpenCodeReasoning-II and Llama-Nemotron synthetic-science question sources, plus unchanged instruction-following/tool-calling data; later reproduction documentation also identifies reused math tool-integrated-reasoning and GenSelect data. Current training documentation specifies standard Qwen2.5 base models, conflicting with launch cards that named Instruct parents. NVIDIA reports 32B pass@1 scores of 89.2 on AIME 2024, 84.0 on AIME 2025, 73.1 on GPQA and 70.2 on LiveCodeBench v6 (August 2024-May 2025). The published recipe averages 64 samples with a 65,536-token generation budget and temperature 0.6 for these benchmarks; these are not best-of-64 scores. Heavy GenSelect inference separately generates and selects among multiple solutions, reporting 93.3 on AIME 2025 and 75.3 on LiveCodeBench for 32B, at additional inference cost. ChatML-style text templates are supplied; no Qwen3 thinking/non-thinking switch is documented. NVIDIA cautions that these are not general-purpose chat assistants, and its later training documentation reports tool-integrated math inference generally performs worse than chain-of-thought. Launch cards declare 64K context, whereas configs permit 131,072 positions and later cards describe evaluation with up to 64K output tokens; these are different limits, not proof of a validated 128K context. July 16 is NVIDIA's declared Hugging Face release date; explicit official download availability is verified by July 18, but the exact first public unlock remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/nvidia/OpenReasoning-Nemotron-1.5B",
        "https://huggingface.co/nvidia/OpenReasoning-Nemotron-7B",
        "https://huggingface.co/nvidia/OpenReasoning-Nemotron-14B",
        "https://huggingface.co/nvidia/OpenReasoning-Nemotron-32B",
        "https://huggingface.co/blog/nvidia/openreasoning-nemotron",
        "https://github.com/NVIDIA-NeMo/Skills/tree/main/docs/releases/openreasoning"
      ],
      "page_url": "https://nlp.com.ai/models/openreasoning-nemotron/",
      "record_updated_at": "2026-10-11T08:37:45.315843+00:00"
    },
    {
      "slug": "voxtral",
      "source_file": "2025-07-15-voxtral.json",
      "release_date": "2025-07-15",
      "date": "July 15, 2025",
      "name": "Voxtral",
      "developer": "Mistral AI (France)",
      "parameters": "Mini: 4,674,351,104; Small: 24,259,880,960 learned parameters, including audio encoder and adapter (nominal 3B/24B text-backbone labels; dense)",
      "context_window": "32K (32,768 tokens supported for both; Transformers text configs specify 131,072, not a verified Voxtral operating limit)",
      "license": "Apache 2.0",
      "key_features": "Original Voxtral Mini 3B 2507 and Small 24B 2507 are dense audio-and-text-input, text-output chat models, retaining text-only generation alongside transcription, speech translation, audio Q&A and summarization; they do not generate audio or include a vision encoder. Mini uses Ministral 3B; the technical report identifies Small as Mistral Small 3.1-derived, although its HF ancestry metadata points to Small 3 Base 2501. A Whisper large-v3-based 32-layer audio encoder processes 16 kHz speech in independently attended, padded 30-second chunks. A two-layer GELU adapter downsamples embeddings 4x to 12.5 Hz before the autoregressive language decoder. Both support 32K-token conversations, with advertised audio durations up to 30 minutes for transcription or 40 minutes for understanding; these durations are not separate token-context sizes. Pretraining balances audio-to-text repetition and cross-modal continuation, with an adapter-only warm-up and text data to preserve language capabilities. SFT combines real speech and synthetic QA, summarization, translation and spoken instruction/tool examples. The released Mini is Online-DPO-aligned; the released Small remains SFT because its preference-aligned experiments slightly regressed ASR. Automatic language detection and English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian support are documented; quality varies substantially by language. The report gives Small 6.31% English short-form macro-average WER versus Mini 6.79% and Gemini 2.5 Flash 8.04%; on FLEURS English, Small/Mini/Whisper large-v3 score 3.35/3.77/4.00% WER, but on Arabic Mini reaches 25.40% versus Whisper 15.44%. Small scores 86.6% on the internal audio-understanding benchmark versus GPT-4o mini Audio 80.0% and Gemini 88.6%, and leads the eight reported FLEURS translation directions. These are task-specific evaluations: ASR comparisons use transcription models, understanding includes filtered speech-synthesized text benchmarks and an LLM-judged internal set, not interchangeable text or general-audio scores. Voice function calling is advertised and Small provides an experimental tool-calling example; launch Mini instructions were inconsistent about implementation support. Multiple audio clips and multi-turn audio are supported; launch instructions do not support system prompts. Official BF16/FP16 GPU-memory estimates are approximately 9.5 GB for Mini and 55 GB for Small, not guaranteed consumer-edge or quantized requirements. The hosted-only original Mini Transcribe optimization and later Voxtral Realtime/2602 models are not included.",
      "paper_url": "https://arxiv.org/abs/2507.13264",
      "links": [
        "https://mistral.ai/news/voxtral/",
        "https://huggingface.co/mistralai/Voxtral-Mini-3B-2507",
        "https://huggingface.co/mistralai/Voxtral-Small-24B-2507",
        "https://docs.mistral.ai/models/voxtral-small-25-07",
        "https://docs.mistral.ai/models/voxtral-mini-transcribe-25-07"
      ],
      "page_url": "https://nlp.com.ai/models/voxtral/",
      "record_updated_at": "2026-10-11T08:37:45.313401+00:00"
    },
    {
      "slug": "ming-lite-omni-v15",
      "source_file": "2025-07-15-ming-lite-omni-v1-5.json",
      "release_date": "2025-07-15",
      "date": "July 15, 2025",
      "name": "Ming-Lite-Omni v1.5",
      "developer": "inclusionAI / Ant Group (China)",
      "parameters": "20.3B total, 3B active in the MoE section",
      "context_window": "32,768 tokens",
      "license": "Hugging Face metadata identifies MIT, but the repository has no root LICENSE file. The packaged talker includes Apache-2.0 terms, while the predecessor's bundled connector carried the non-commercial Qwen Research License; applicability of that earlier component license to the v1.5 package remains unresolved.",
      "key_features": "An any-to-any multimodal model built on Ling-lite-1.5 that accepts text, images, video and audio and can generate text, speech and images. Version 1.5 upgrades the earlier Ming-Lite-Omni with MRoPE-based spatiotemporal encoding and curriculum learning for long-video understanding; dual-branch image generation and editing with identity- and scene-consistency objectives; a new audio decoder and BPE-based speech representation for real-time speech synthesis; and expanded structured-text, product, visual and speech training data. The official model card reports a 20.3B-parameter system with 3B active parameters in its MoE section and reports a 69.83 average on VideoMME, alongside improvements in document understanding, OCR, chart reasoning and multimodal generation. The shipped language-model configuration uses 32,768 maximum position embeddings. Downloadable BF16 weights are provided through both Hugging Face and ModelScope, with a separately released FP8 package excluded from the headline parameter count.",
      "paper_url": "https://arxiv.org/abs/2506.09344",
      "links": [
        "https://huggingface.co/inclusionAI/Ming-Lite-Omni-1.5",
        "https://www.modelscope.cn/models/inclusionAI/Ming-Lite-Omni-1.5",
        "https://github.com/inclusionAI/Ming/tree/v1.5",
        "https://inclusionai.github.io/blog/ming-lite-omni-1_5/",
        "https://lucaria-academy.github.io/Ming-Omni/"
      ],
      "page_url": "https://nlp.com.ai/models/ming-lite-omni-v15/",
      "record_updated_at": "2026-10-11T08:37:45.310771+00:00"
    },
    {
      "slug": "kimi-k2",
      "source_file": "2025-07-11-kimi-k2.json",
      "release_date": "2025-07-11",
      "date": "July 11, 2025",
      "name": "Kimi K2",
      "developer": "Moonshot AI (China)",
      "parameters": "1T total / 32B active per token (sparse MoE; paper: 1.04T / 32.6B)",
      "context_window": "128K (131,072 tokens)",
      "license": "Modified MIT License (retain copyright/permission notices; prominently display \"Kimi K2\" in a commercial product/service UI if it exceeds 100 million monthly active users OR US$20 million monthly revenue)",
      "key_features": "Original text-only Kimi-K2-Base and Kimi-K2-Instruct release: Base supports research and custom fine-tuning; Instruct is the chat/tool-use model, described as reflex-grade without a dedicated long-thinking mode. The 61-layer decoder uses Multi-head Latent Attention (MLA), one initial dense layer, and 384 routed experts with eight selected per token plus one shared expert. Pretrained on 15.5T tokens with MuonClip, which augments Muon with post-update query/key weight clipping to control exploding attention logits; Moonshot reports zero loss spikes. Post-training combines supervised instruction tuning, scalable synthetic tool-use trajectories in simulated and real environments, and joint reinforcement learning with verifiable rewards and self-critique rubric feedback. Official non-thinking Instruct results include 53.7% LiveCodeBench v6 pass@1, 49.5% AIME 2025 avg@64, and 65.8% SWE-bench Verified with bash/editor tools and a single attempt; 71.6% uses multiple sampled attempts and an internal scoring model, not the same inference budget. Strong agentic coding and tool use are conditional on an appropriate execution framework: the report finds one-shot complete-project prompting less reliable, and warns that hard reasoning or unclear tool definitions can cause excessive output, truncation or incomplete calls; unnecessarily enabling tools can also reduce performance. Original weights use block-FP8 packaging, lowering storage relative to BF16 but retaining a trillion-parameter deployment footprint; 32B active does not mean 32B resident weights. The July release has 128K context, not the later 0905 variant's 256K; K2-Instruct-0905 and K2-Thinking are separate later releases.",
      "paper_url": "https://arxiv.org/abs/2507.20534",
      "links": [
        "https://github.com/MoonshotAI/Kimi-K2",
        "https://huggingface.co/moonshotai/Kimi-K2-Base",
        "https://huggingface.co/moonshotai/Kimi-K2-Instruct",
        "https://modelscope.cn/models/moonshotai/Kimi-K2-Base",
        "https://modelscope.cn/models/moonshotai/Kimi-K2-Instruct"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-k2/",
      "record_updated_at": "2026-10-11T08:37:45.308130+00:00"
    },
    {
      "slug": "lfm2",
      "source_file": "2025-07-10-lfm2.json",
      "release_date": "2025-07-10",
      "date": "July 10, 2025",
      "name": "LFM2",
      "developer": "Liquid AI (USA)",
      "parameters": "350M, 700M, 1.2B dense (354,483,968; 742,489,344; 1,170,340,608 unique parameters)",
      "context_window": "32,768 tokens",
      "license": "LFM Open License v1.0 (custom, not Apache 2.0; launch guidance permits commercial use below US$10M annual revenue, with a separate commercial license above the threshold)",
      "key_features": "Original release of three post-trained, text-only edge models, not the later 2.6B, MoE, vision, audio, or LFM2.5 releases. Each uses 16 layers: 10 double-gated, kernel-size-3 short convolutions and 6 grouped-query attention blocks with SwiGLU and RMSNorm; dense hybrid architecture with a short convolution cache, not a neural-ODE or Mamba model. Tied input/output embeddings and a 65,536-token vocabulary; ChatML-like prompts and function-calling special tokens, with no documented thinking-mode switch. Supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. Launch sources report 10T-token training on approximately 75% English, 20% multilingual, and 5% code, with LFM1-7B teacher distillation, context extension to 32K, large-scale SFT (50% downstream tasks / 50% general domains), length-normalized DPO on offline and semi-online preference pairs, and checkpoint merging; the later technical report describes an additional 1T long-context phase, so token-budget accounting differs. No original on-policy RL stage is established. Designed for local assistants, data extraction, RAG, creative writing, multi-turn dialogue, and function calling; Liquid recommends task-specific fine-tuning and advises against knowledge-intensive or programming tasks. Launch internal evaluations report 74.89 IFEval and 58.3 GSM8K for 1.2B, versus 73.98 and 51.4 for Qwen3-1.7B in non-reasoning mode; Qwen3 remains stronger on MMLU and MGSM, so superiority is not universal. Liquid claims 3x training efficiency versus LFM1 and up to 2x CPU prefill/decode speed versus Qwen3, measured with quantized ExecuTorch/llama.cpp on Samsung Galaxy S24 Ultra and AMD Ryzen HX370; results depend on hardware, runtime, quantization, and sequence length, not a universal CPU/GPU/NPU or Apple-device speed guarantee. Supported context is 32K despite a 128,000 positional config ceiling; no separate maximum-output budget is published.",
      "paper_url": "https://arxiv.org/abs/2511.23404",
      "links": [
        "https://huggingface.co/LiquidAI/LFM2-350M",
        "https://huggingface.co/LiquidAI/LFM2-700M",
        "https://huggingface.co/LiquidAI/LFM2-1.2B",
        "https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models",
        "https://www.liquid.ai/press/liquid-ai-releases-worlds-fastest-and-best-performing-open-source-small-foundation-models",
        "https://huggingface.co/LiquidAI/LFM2-1.2B/blob/1f6a681771e28ba0ea43cb1538ba1b7b66bf865a/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/lfm2/",
      "record_updated_at": "2026-10-11T08:37:45.306398+00:00"
    },
    {
      "slug": "t5gemma",
      "source_file": "2025-07-09-t5gemma.json",
      "release_date": "2025-07-09",
      "date": "July 9, 2025",
      "name": "T5Gemma",
      "developer": "Google (USA)",
      "parameters": "Dense encoder-decoder totals: S-S 312,517,632; B-B 591,490,560; L-L 1,241,761,792; ML-ML 2,200,345,344; XL-XL 3,766,980,608; 2B-2B 5,596,853,760; 9B-2B 12,292,375,296; 9B-9B 20,333,401,088. Size-pair labels name encoder/decoder configurations, not summed checkpoint totals.",
      "context_window": "Unknown (supported checkpoint maximum not verified; mirrors configure 8,192 positions separately for encoder and decoder)",
      "license": "Custom Gemma Terms of Use plus Gemma Prohibited Use Policy (all pretrained and IT weights); not Apache 2.0. Implementation code is Apache 2.0.",
      "key_features": "Original text-only, primarily English T5Gemma research family: eight encoder-decoder configurations, each released as UL2 and PrefixLM pretrained checkpoints plus their instruction-tuned counterparts (32 checkpoints); excludes December T5Gemma 2. Uses a bidirectional encoder, causal autoregressive decoder and separate global cross-attention, with Gemma 2-style alternating local/global self-attention, RoPE, gated GELU and RMSNorm; Gemma-sized variants use grouped-query attention. Encoder and decoder embeddings are separate; decoder input/output embeddings are tied. The 2B-2B, 9B-2B and 9B-9B models adapt pretrained Gemma 2 weights with 2T further tokens; Small/Base/Large/Medium-Large/XL first receive 2T decoder-only pretraining under the Gemma 2 recipe and then 2T encoder-decoder adaptation. PrefixLM uses prefix-to-continuation learning with knowledge distillation; UL2 combines denoising objectives and favors contextual representations. Balanced models initialize cross-attention from self-attention; the unbalanced 9B-2B model warms up newly initialized cross-attention before tuning all weights. Public pretrained variants can generate text via AutoModelForSeq2SeqLM but are not instruction-following chat models; released IT variants support Gemma turn formatting. The paper studies supervised instruction tuning and RLHF, but the public cards do not establish the exact post-training stage of every IT checkpoint. Supports QA, summarization and reasoning, with encoders useful for classification after task-specific fine-tuning. Google reports an improved quality/inference-efficiency frontier; paper PrefixLM RLHF experiments score 61.5 vs 49.8 on MMLU and 70.7 vs 58.0 on GSM8K for 2B-2B versus Gemma 2 2B. These are experimental post-training results, not independently verified scores for every released IT variant, and compare similar inference budgets rather than equal parameter counts or total training compute. Benefits are task-dependent: the paper reports weaker pretrained ARC-C and some instruction-tuned translation results than decoder-only comparators. Primarily English evaluation, factual errors, bias and harmful output remain limitations; translation use cases do not imply broad multilingual certification. Encoder and decoder position limits are separate, not a single 16K context; training sequence lengths are objective-specific and do not establish validated long-context support.",
      "paper_url": "https://arxiv.org/abs/2504.06225",
      "links": [
        "https://developers.googleblog.com/t5gemma/",
        "https://huggingface.co/collections/google/t5gemma-686ba262fe290b881d21ec86",
        "https://huggingface.co/google/t5gemma-2b-2b-ul2",
        "https://huggingface.co/google/t5gemma-2b-2b-prefixlm-it",
        "https://www.kaggle.com/models/google/t5gemma",
        "https://github.com/google-deepmind/gemma/tree/main/gemma/research/t5gemma",
        "https://ai.google.dev/gemma/docs/releases",
        "https://ai.google.dev/gemma/terms",
        "https://ai.google.dev/gemma/prohibited_use_policy"
      ],
      "page_url": "https://nlp.com.ai/models/t5gemma/",
      "record_updated_at": "2026-10-11T08:37:45.304061+00:00"
    },
    {
      "slug": "smollm3-3b",
      "source_file": "2025-07-08-smollm3-3b.json",
      "release_date": "2025-07-08",
      "date": "July 8, 2025",
      "name": "SmolLM3-3B",
      "developer": "Hugging Face (USA)",
      "parameters": "3B dense (3,075,098,624 unique parameters per checkpoint; tied input/output embeddings)",
      "context_window": "65,536 tokens trained/configured; up to 131,072 tokens with 2× YaRN extrapolation",
      "license": "Apache 2.0",
      "key_features": "Original text-only SmolLM3 release groups SmolLM3-3B-Base and the post-trained SmolLM3-3B checkpoint; reasoning and non-reasoning are runtime modes of the latter, not separate weights. A Llama-style dense decoder with 36 layers, 2,048 hidden width, SwiGLU, 16 query heads and 4 KV heads uses GQA and interleaves three RoPE layers with one NoPE layer (27 RoPE/9 NoPE), reducing KV-cache requirements and improving long-context handling. Native target languages are English, French, Spanish, German, Italian and Portuguese; Arabic, Chinese and Russian received less training data. Hugging Face reports approximately 11.2T pretraining tokens with a three-stage web/code/math curriculum: 85/12/3%, then 75/15/10%, then 63/24/13%. Sources include FineWeb-Edu, DCLM, FineWeb2/FineWeb2-HQ, The Stack v2 and StarCoder2-related code, notebooks, GitHub issues, StackExchange, Stack-Edu, FineMath/InfiWebMath, MegaMath and late-stage OpenMathReasoning. Educational-quality filtering and increasingly selective math/code subsets prioritize training quality. Training used nanotron/datatrove, intra-document attention masking, a warmup-stable-decay schedule and 384 H100 GPUs for 24 days. Long-context adaptation added two 50B-token stages at 32K and 64K; 128K is inference extrapolation, not the training length or default positional setting. The post-trained model adds roughly 140B reasoning-midtraining tokens from OpenThoughts3 and NVIDIA Llama-Nemotron reasoning traces, dual-mode SFT on 1.8B tokens for four epochs including Qwen3-32B-generated traces, and off-policy Anchored Preference Optimization (a DPO variant), not a reported on-policy RL stage. The final checkpoint merges an APO checkpoint soup with a long-context midtraining checkpoint at 0.9/0.1 to recover long-context performance. Extended thinking is enabled by default; system /think and /no_think flags override enable_thinking, while /system_override removes the metadata header. XML/JSON and Python-style tool calling are supported. Publisher evaluations report non-thinking IFEval 76.7 and BFCL 92.3; thinking raises AIME 2025 from 9.3 to 36.7 and LiveCodeBench v4 from 15.2 to 30.0, but lowers IFEval to 71.2 and BFCL to 88.8. These are publisher-protocol results, not universal BFCL leaderboard or deployment guarantees; the tool-call comparator includes a specialized finetune. Competitive with other 3B models and some 4B results, but Qwen3-4B leads many reasoning benchmarks and base coding/long-context results are not uniformly best. The publisher warns of factual errors, inconsistent reasoning and training-data biases. July 8 is supported public availability-by for both original checkpoints; exact first public unlock remains unresolved.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/HuggingFaceTB/SmolLM3-3B",
        "https://huggingface.co/HuggingFaceTB/SmolLM3-3B-Base",
        "https://huggingface.co/blog/smollm3",
        "https://github.com/huggingface/smollm",
        "https://huggingface.co/ggml-org/SmolLM3-3B-GGUF"
      ],
      "page_url": "https://nlp.com.ai/models/smollm3-3b/",
      "record_updated_at": "2026-10-11T08:37:45.301679+00:00"
    },
    {
      "slug": "mistral-small-32",
      "source_file": "2025-06-20-mistral-small-3-2.json",
      "release_date": "2025-06-20",
      "date": "June 20, 2025",
      "name": "Mistral Small 3.2",
      "developer": "Mistral AI (France)",
      "parameters": "24B dense (24.011B total: 23.572B text + 403M vision + 35.7M multimodal connector)",
      "context_window": "128K (131,072 tokens)",
      "license": "Apache 2.0",
      "key_features": "Original Mistral-Small-3.2-24B-Instruct-2506: a targeted instruction-tuned update to Small 3.1, with the same dense multimodal Mistral3 architecture and Small-3.1-24B-Base-2503 ancestry. Combines a 40-layer Mistral decoder (5,120 hidden width, 32 query/8 KV heads) with a 24-layer Pixtral vision encoder and patch-merging connector; accepts text and images and generates text, without audio support. Supports multilingual chat, image understanding, document Q&A, JSON output and function calling. Native 128K context; Tekken v11 tokenizer, 14-pixel vision patches with 2x2 spatial merging and variable image-token usage. Mistral reports better precise instruction following, fewer repetitive/infinite generations and a more robust tool-call template. Its Small 3.1-to-3.2 comparison reports internal instruction accuracy 82.75% to 84.78%, Wildbench v2 55.6% to 65.33%, Arena Hard v2 19.56% to 43.1%, and infinite-generation errors 2.11% to 1.29%. STEM results include MMLU Pro (5-shot CoT) 69.06%, MATH 69.42%, MBPP Plus pass@5 78.33% and HumanEval Plus pass@5 92.90%; these are publisher evaluations, not single-attempt coding scores. Improvements are not uniform: MMLU and GPQA Main slightly decline, MMMU falls from 64.00% to 62.50% and MathVista from 68.91% to 67.09%, while ChartQA and DocVQA improve; SimpleQA remains only 12.10% total accuracy. No release-specific training/RL recipe is disclosed. Official examples recommend temperature 0.15 and the supplied system prompt, vLLM >=0.9.1 with mistral-common >=1.6.2 and the Mistral tool-call parser, or Transformers multimodal generation; BF16/FP16 serving requires approximately 55 GB GPU RAM. Only the original Instruct checkpoint is included; no publicly verified official Small 3.2 Base or GGUF sibling, and Small 3.1/later versions are separate releases.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506",
        "https://docs.mistral.ai/models/mistral-small-3-2-25-06"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-small-32/",
      "record_updated_at": "2026-10-11T08:37:45.299270+00:00"
    },
    {
      "slug": "minimax-m1",
      "source_file": "2025-06-16-minimax-m1.json",
      "release_date": "2025-06-16",
      "date": "June 16, 2025",
      "name": "MiniMax-M1",
      "developer": "MiniMax (China)",
      "parameters": "456B total; 45.9B active per token (MoE; active count excludes embedding/output matrices)",
      "context_window": "1M tokens input; 40K / 80K maximum reasoning output by checkpoint",
      "license": "Apache 2.0 (both released M1 checkpoints; Copyright 2025 MiniMax)",
      "key_features": "Text-only hybrid-attention reasoning family based on MiniMax-Text-01, releasing MiniMax-M1-40k and MiniMax-M1-80k together; the 40K checkpoint is an intermediate stage of the 80K training run, not a runtime toggle. Its 80-layer Transformer combines seven Lightning Attention layers followed by one softmax-attention layer, repeated ten times, with 32 routed experts and two selected per token; it is not a Mamba/state-space model. Both checkpoints have 456,089,655,296 stored tensor parameters; the reported 45.9B active count excludes the untied input embedding and output head. Publisher-supported 1M-token input context is distinct from 40K/80K output budgets and the config positional ceiling of 10,240,000. Training adds 7.5T tokens of continual pretraining, cold-start long-chain-of-thought SFT, and diverse reinforcement learning with mathematical/code verifiers, execution-based software-engineering sandboxes and generative reward models. CISPO clips stop-gradient importance-sampling weights while retaining token gradients and GRPO-style group-relative advantages, rather than using PPO/GRPO token-update clipping. The publisher reports the full RL phase took three weeks on 512 H800 GPUs at $534,700 rental cost; this is not the cost of all pretraining and SFT. Supports long reasoning and structured function calls for tool-using agents. Publisher evaluations use temperature 1.0/top-p 0.95: M1-80k scores 86.0% AIME 2024 and 65.0% LiveCodeBench (average pass rates over 32 and 16 samples respectively, not best-of-N), 8.4% HLE on the text-only subset without tools, and 56.0% SWE-bench Verified using a modified Agentless scaffold on 486/500 tasks. TAU-bench airline/retail scores are 62.0%/63.5% for 80k and 60.0%/67.8% for 40k, with GPT-4.1 as simulated user, standard benchmark tools and at most 40 interaction steps. Strengths emphasize long-context retrieval, software engineering and tool use; it trails DeepSeek-R1-0528 on competitive math/coding and SWE-bench, and does not outperform original R1 on GPQA or SimpleQA. June 16 is supported by the official dated statement that complete weights were available; exact first public unlock remains unresolved, and June 17 API/publication references are not established checkpoint-release dates.",
      "paper_url": "https://arxiv.org/abs/2506.13585",
      "links": [
        "https://huggingface.co/MiniMaxAI/MiniMax-M1-40k",
        "https://huggingface.co/MiniMaxAI/MiniMax-M1-80k",
        "https://modelscope.cn/models/MiniMax/MiniMax-M1-40k",
        "https://modelscope.cn/models/MiniMax/MiniMax-M1-80k",
        "https://github.com/MiniMax-AI/MiniMax-M1",
        "https://www.minimax.io/news/minimaxm1"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-m1/",
      "record_updated_at": "2026-10-11T08:37:45.296658+00:00"
    },
    {
      "slug": "magistral-small-10-2506",
      "source_file": "2025-06-10-magistral-small.json",
      "release_date": "2025-06-10",
      "date": "June 10, 2025",
      "name": "Magistral Small 1.0 (2506)",
      "developer": "Mistral AI (France)",
      "parameters": "24B (23,572,403,200; dense)",
      "context_window": "128K declared; 40K recommended (40,960 configured currently; 32,768 in launch HF config)",
      "license": "Apache 2.0",
      "key_features": "Mistral AI’s first open-weight reasoning checkpoint: a text-only, 40-layer dense decoder Transformer with grouped-query attention (32 query heads, 8 KV heads), derived from Mistral Small 3.1 Instruct 2503 without the parent’s vision encoder. Trained by supervised distillation of Magistral Medium reasoning traces followed by modified GRPO reinforcement learning on verifiable math and coding tasks; not a pure-RL/no-SFT release. The report describes four SFT epochs with 10% general instruction data, then RL. Produces visible <think> reasoning and a final answer, with multilingual reasoning encouraged to match the user’s language; the card lists dozens of supported languages, not merely bilingual reasoning. Publisher evaluations report AIME 2024/2025 pass@1 of 70.68%/62.76%, GPQA Diamond 68.18%, and LiveCodeBench v5 55.84%; AIME majority voting over 64 samples reaches 83.3%/76.7%, distinct from single-sample accuracy. Report evaluations allow up to 40K generated tokens for AIME/code and 32K for other tasks; the published protocol does not specify external-tool assistance, and scores are not fixed-latency guarantees. Strengths include multi-step mathematical, coding and multilingual reasoning, with local deployment and official same-release GGUF packaging; fitting a single RTX 4090 or 32GB MacBook requires quantization. Although the card declares 128K context, it warns that performance may degrade beyond 40K and recommends a 40K model-length cap. Published Small results also remain limited on text-only Humanity’s Last Exam (6.4%). The hosted Medium/Le Chat Flash Answers claim of up to 10x throughput does not describe local Small inference. Closed Magistral Medium and later 2507/2509 checkpoints are excluded.",
      "paper_url": "https://arxiv.org/abs/2506.10910",
      "links": [
        "https://huggingface.co/mistralai/Magistral-Small-2506",
        "https://huggingface.co/mistralai/Magistral-Small-2506_gguf",
        "https://mistral.ai/news/magistral/",
        "https://docs.mistral.ai/models/magistral-small-1-0-25-06"
      ],
      "page_url": "https://nlp.com.ai/models/magistral-small-10-2506/",
      "record_updated_at": "2026-10-11T08:37:45.294436+00:00"
    },
    {
      "slug": "deepseek-r1-0528",
      "source_file": "2025-05-28-deepseek-r1-0528.json",
      "release_date": "2025-05-28",
      "date": "May 28, 2025",
      "name": "DeepSeek-R1-0528",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B main MoE / 37B activated; approximately 685B packaged including the auxiliary MTP layer",
      "context_window": "128K supported; 163,840 positions configured with YaRN",
      "license": "MIT (code and model weights; commercial use and distillation permitted)",
      "key_features": "Text-only reasoning update using the same December 2024 DeepSeek-V3-Base as the original R1, with more post-training compute and algorithmic optimizations rather than a newly pretrained base. Its 61-layer decoder combines Multi-head Latent Attention (MLA) with DeepSeekMoE: 256 routed experts, eight selected per token, and one shared expert per MoE layer; the released FP8 checkpoint also packages one auxiliary multi-token-prediction (MTP) layer. DeepSeek reports stronger mathematics, coding and general reasoning, including AIME 2025 87.5% versus 70.0%, GPQA-Diamond 81.0%, LiveCodeBench (August 2024–May 2025) 73.3%, and text-only Humanity's Last Exam 17.7% pass@1. Sampling-based results estimate pass@1 from 16 responses per question at temperature 0.6/top-p 0.95 with a 64K generation ceiling, not best-of-16 accuracy; SWE-bench Verified 57.6% uses Agentless and is a separate tool-assisted evaluation. The update supports system prompts, function calling outside the thinking phase, and JSON output, and improves front-end code generation, role-playing and long-form writing. DeepSeek claims 45–50% fewer hallucinations in rewriting, summarization and reading comprehension, but SimpleQA correctness declines from 30.1% to 27.8%; the improvement claim is not universal factual reliability. Deeper reasoning is substantially more verbose: average AIME reasoning grows from 12K to 23K tokens per question. Open weights support 128K context, distinct from the original hosted service's 64K limit. The related Qwen3-8B dense distillation is not grouped here because its public download evidence is May 29 rather than confirmed May 28; the six January R1 distillations are also excluded. The linked paper describes the original January R1 training pipeline, not a dedicated technical report for this May update.",
      "paper_url": "https://arxiv.org/abs/2501.12948v1",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-0528",
        "https://api-docs.deepseek.com/zh-cn/news/news250528",
        "https://api-docs.deepseek.com/news/news250528",
        "https://www.deepseek.com/news/r1-0528/",
        "https://github.com/deepseek-ai/DeepSeek-R1"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-r1-0528/",
      "record_updated_at": "2026-10-11T08:37:45.292049+00:00"
    },
    {
      "slug": "falcon-h1",
      "source_file": "2025-05-21-falcon-h1.json",
      "release_date": "2025-05-21",
      "date": "May 21, 2025",
      "name": "Falcon-H1",
      "developer": "Technology Innovation Institute (TII) (UAE)",
      "parameters": "Six dense configurations: 0.5B (~0.521B), 1.5B (~1.555B), 1.5B-Deep (~1.555B), 3B (~3.149B), 7B (~7.586B), and 34B (~33.643B); each released as Base and Instruct",
      "context_window": "16K (16,384 tokens): 0.5B; 128K (131,072 tokens): 1.5B, 1.5B-Deep, 3B; 256K (262,144 tokens): 7B, 34B",
      "license": "Custom TII Falcon / Falcon-LLM License (Apache-2.0-based, not Apache 2.0): all six Base/Instruct sizes; incorporated Acceptable Use Policy, downstream enforceable use restrictions, license/notice retention and derivative-publication attribution",
      "key_features": "Original text-only causal language-model family with parallel hybrid mixer blocks: grouped-query Transformer attention and Mamba-2 state-space heads run together within every block, their outputs are concatenated and projected, then processed by a dense feed-forward network; not MoE or alternating Transformer/Mamba layers. Six Base/Instruct pairs plus official GPTQ 4-bit/8-bit and GGUF packaging were offered in the launch release. The 1.5B-Deep configuration uses 66 narrower layers versus 24 wider layers in the standard 1.5B, trading throughput for quality. Customized maximal update parametrization (muP), high-quality STEM/code data, reuse of curated data, and early exposure to difficult/long-context examples underpin training; the technical report gives pretraining budgets of 2.5T tokens for 0.5B and 3B, 3T for both 1.5B configurations, approximately 12T for 7B and 18T for 34B. Instruct models use supervised fine-tuning followed by direct preference optimization, without reasoning-specific fine-tuning or a documented launch thinking-mode toggle. The 0.5B model is English-only; larger configurations support 18 trained languages including Arabic, while 100+ languages is a tokenizer scalability claim rather than demonstrated native coverage. Context support is size-specific, not uniformly 256K: the original 0.5B-Instruct config advertised a 131,072 positional ceiling, subsequently corrected to 16,384; the report specifies 16K and omits its 128K SFT stage. TII reports strong math, science, coding, instruction-following and multilingual results, with 1.5B-Deep competitive with selected 7B-10B models and 34B competitive with selected 70B-class models; Qwen3 comparisons disable thinking. In the reported long-context evaluation, 34B-Instruct leads RAG at 131K but trails Qwen3-32B/Llama3.3-70B on some pure-recall and long-document-QA tasks. Reported up-to-4x prefill and 8x generation throughput gains versus Qwen2.5-32B apply to long sequences and the tested vLLM implementations; Transformers are slightly faster at short contexts. Launch inference used Transformers, vLLM and a custom llama.cpp fork, not all later upstream integrations. Function-calling training/template support exists, but the developer acknowledged incomplete launch templates and corrected the remaining models by May 31; uniform launch-ready tool use is not established. Later Falcon-H1R, H1-Arabic and H1-Tiny releases are excluded. May 21 is the provisional public-weight date supported by the official download-linked release; exact first public unlock remains unresolved. The separate blog displays May 20 UTC, but its original source date explicitly says todo update and was committed May 21.",
      "paper_url": "https://arxiv.org/abs/2507.22448",
      "links": [
        "https://huggingface.co/collections/tiiuae/falcon-h1-6819f2795bc406da60fab8df",
        "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Instruct",
        "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Instruct",
        "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Deep-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Deep-Instruct",
        "https://huggingface.co/tiiuae/Falcon-H1-3B-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-3B-Instruct",
        "https://huggingface.co/tiiuae/Falcon-H1-7B-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-7B-Instruct",
        "https://huggingface.co/tiiuae/Falcon-H1-34B-Base",
        "https://huggingface.co/tiiuae/Falcon-H1-34B-Instruct",
        "https://huggingface.co/blog/tiiuae/falcon-h1",
        "https://github.com/tiiuae/Falcon-H1",
        "https://falcon-lm.github.io/blog/falcon-h1/",
        "https://tiiuae.github.io/Falcon-H1/",
        "https://falconllm.tii.ae/falcon-terms-and-conditions.html",
        "https://falconllm.tii.ae/falcon-acceptable-use-policy.html",
        "https://www.tii.ae/news/middle-easts-leading-ai-powerhouse-tii-launches-two-new-ai-models-falcon-arabic-first-arabic"
      ],
      "page_url": "https://nlp.com.ai/models/falcon-h1/",
      "record_updated_at": "2026-10-11T08:37:45.289761+00:00"
    },
    {
      "slug": "devstral-small-2505-devstral-small-10",
      "source_file": "2025-05-21-devstral.json",
      "release_date": "2025-05-21",
      "date": "May 21, 2025",
      "name": "Devstral Small 2505 (Devstral Small 1.0)",
      "developer": "Mistral AI (France) and All Hands AI (USA)",
      "parameters": "24B dense (23,572,403,200 parameters)",
      "context_window": "128K tokens (131,072 configured)",
      "license": "Apache 2.0",
      "key_features": "Original text-only agentic coding research preview, fine-tuned from the Mistral Small 3.1 family with the vision encoder removed; not the later Devstral Small 2507 or Devstral 2. A 40-layer dense decoder-only Transformer with grouped-query attention (32 query heads, 8 KV heads) and a 131,072-token Tekken vocabulary. The joint technical report describes supervised OpenHands CodeAct trajectories generated in SWE-Gym, unit-test-based patch checks, two stages of increasingly strict trajectory filtering, then additional model rollouts and policy optimization; selected natural-language data retain general language understanding. Designed for tool-driven repository exploration, multi-file edits and real GitHub issue resolution. Publishers report 46.8% SWE-bench Verified resolution with OpenHands on 500 issues; the report specifies bash/file-edit tools, disabled web browsing and an iterative protocol allowing up to three attempts, not standalone no-tool code generation or single-attempt pass@1. Larger open-model comparisons use the same scaffold/protocol; closed-model comparisons use different scaffolds. The first-attempt ablation scores 43.0%, and reducing the iteration limit from 50 to 30 lowers resolution to 36.8%; increasing it to 100 does not improve 46.8%. Official BF16 and GGUF packages support local deployment; the single-RTX-4090/32GB-Mac claim depends on reduced-precision packaging, since BF16 parameter data alone occupy approximately 47.1GB. No image input after vision removal. Release-date review remains open: Mistral explicitly offers downloads starting May 21, while the collaborator page is dated May 20; exact first public availability is unresolved.",
      "paper_url": "https://arxiv.org/abs/2509.25193",
      "links": [
        "https://mistral.ai/news/devstral",
        "https://huggingface.co/mistralai/Devstral-Small-2505",
        "https://huggingface.co/mistralai/Devstral-Small-2505_gguf",
        "https://www.openhands.dev/blog/devstral-a-new-state-of-the-art-open-model-for-coding-agents",
        "https://docs.mistral.ai/models/devstral-small-1-0-25-05"
      ],
      "page_url": "https://nlp.com.ai/models/devstral-small-2505-devstral-small-10/",
      "record_updated_at": "2026-10-11T08:37:45.287829+00:00"
    },
    {
      "slug": "harbinger-24b",
      "source_file": "2025-05-16-harbinger-24b.json",
      "release_date": "2025-05-16",
      "date": "May 16, 2025",
      "name": "Harbinger-24B",
      "developer": "LatitudeGames (USA)",
      "parameters": "24B (23,572,403,200 parameters; dense)",
      "context_window": "131,072 tokens configured; finetune-specific supported maximum unverified",
      "license": "Apache 2.0",
      "key_features": "Text-only storytelling and roleplay finetune of Mistral-Small-3.1-24B-Instruct-2503, developed with Gryphe Padar; the released checkpoint uses a 40-layer Mistral causal decoder with grouped-query attention and no vision encoder. Trained in two stated stages: supervised finetuning on balanced, rewritten multi-turn Wayfarer-style adventures and general roleplay, plus a small single-turn instruction dataset; then DPO using reward-model-derived user-preference data. Latitude claims improved instruction following, mid-sequence continuation, long-run narrative coherence, more consistent characters, and fewer clichés and repetitive patterns while preserving high-stakes, unforgiving stories. ChatML was used throughout training. The parent supports 128K context, and Harbinger configures 131,072 positions, but the publisher provides no finetune-specific maximum-context validation. Training exclusively used second-person present-tense narrative data; other styles may produce suboptimal results. Recommended baseline sampling is temperature 0.8, repetition penalty 1.05, and min_p 0.025. Official BF16 and GGUF weights are downloadable without an AI Dungeon subscription; the subscription requirement applies to the hosted Harbinger service. Release date remains provisional pending verification of first public weight availability.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Harbinger-24B",
        "https://huggingface.co/LatitudeGames/Harbinger-24B-GGUF",
        "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503",
        "https://blog.latitude.io/all-posts/synthetic-data-preference-optimization-and-reward-models",
        "https://blog.latitude.io/all-posts/your-saga-begins-meet-your-new-ai-models"
      ],
      "page_url": "https://nlp.com.ai/models/harbinger-24b/",
      "record_updated_at": "2026-10-11T08:37:45.286109+00:00"
    },
    {
      "slug": "muse-12b",
      "source_file": "2025-05-13-muse-12b.json",
      "release_date": "2025-05-13",
      "date": "May 13, 2025",
      "name": "Muse-12B",
      "developer": "LatitudeGames (USA)",
      "parameters": "12.248B (12,247,802,880; dense)",
      "context_window": "131,072 tokens (configured; finetune support unverified)",
      "license": "Apache 2.0",
      "key_features": "English storytelling and roleplay finetune of Mistral-Nemo-Base-2407, developed by Latitude with Gryphe Padar; dense 40-layer Mistral causal Transformer with grouped-query attention and BF16 weights. Uses ChatML throughout three training stages: supervised fine-tuning on multi-turn text adventures, long emotional narratives and roleplay rewritten to remove common AI clichés, plus a small single-turn instruction dataset; Gutenberg DPO to introduce human writing techniques; and reward-model-ranked user-preference DPO to refine writing while recovering intelligence lost in the Gutenberg stage. Latitude highlights emotional depth, character relationships, genre versatility and narrative coherence over long contexts; Muse is also available free in AI Dungeon. Training is exclusively second-person present-tense narrative, so other styles may be suboptimal; Gutenberg DPO causes verbose responses averaging 1,000+ tokens, which system instructions can constrain. Recommended baseline sampling is temperature 0.8, repetition penalty 1.05 and min_p 0.025. Official GGUF conversions are available. The checkpoint config permits 131,072 positions, but a Muse-specific validated maximum is not published; AI Dungeon’s 4K–32K tier limits are service settings, not the checkpoint limit.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Muse-12B",
        "https://huggingface.co/LatitudeGames/Muse-12B-GGUF",
        "https://latitude.io/news/your-saga-begins-meet-your-new-ai-models",
        "https://latitude.io/news/synthetic-data-preference-optimization-and-reward-models",
        "https://help.aidungeon.com/ai-model-differences"
      ],
      "page_url": "https://nlp.com.ai/models/muse-12b/",
      "record_updated_at": "2026-10-11T08:37:45.283787+00:00"
    },
    {
      "slug": "qwen3-original-april-2025-release",
      "source_file": "2025-04-28-qwen3.json",
      "release_date": "2025-04-28",
      "date": "April 28, 2025",
      "name": "Qwen3 (original April 2025 release)",
      "developer": "Alibaba Cloud / Qwen Team (China)",
      "parameters": "Dense: 0.6B, 1.7B, 4.0B, 8.2B, 14.8B, 32.8B total; MoE: 30.5B total / 3.3B active and 235B total / 22B active (including embeddings)",
      "context_window": "32,768 tokens native for all sizes; 131,072 with explicitly enabled YaRN for post-trained 4B, 8B, 14B, 32B, 30B-A3B and 235B-A22B; 0.6B/1.7B and released Base cards specify 32,768; post-trained configs default to 40,960, while 8B-Base config is 32,768",
      "license": "Apache License 2.0 (all 14 confirmed original post-trained/Base checkpoints)",
      "key_features": "Original text-only, decoder-only Qwen3 family: six dense sizes (0.6B/1.7B/4B/8B/14B/32B) and two mixture-of-experts models (30B-A3B/235B-A22B). All eight have post-trained hybrid reasoning checkpoints without an -Instruct suffix; publicly downloadable Base counterparts are confirmed only for 0.6B, 1.7B, 4B, 8B, 14B and 30B-A3B, not 32B or 235B-A22B. Uses grouped-query attention, QK normalization, bias-free QKV projections, SwiGLU, RoPE and pre-RMSNorm; MoE layers have 128 routed experts, eight active per token and no shared experts. Pre-trained on approximately 36 trillion tokens spanning 119 languages and dialects, including web/document text and synthetic mathematics/code; three stages move from general 4K training through reasoning-focused data to 32K long-context training. Post-training combines long-chain-of-thought cold start, reasoning RL, thinking/non-thinking fusion and general RL; smaller models additionally use strong-to-weak distillation. Hybrid checkpoints default to thinking, with enable_thinking=False as a hard chat-template switch and /think or /no_think as soft turn-level controls; inference-time thinking budgets trade reasoning quality against latency. The Qwen3 Technical Report evaluates the original flagship in thinking mode at 85.7 on AIME 2024, 81.5 on AIME 2025, 70.7 on LiveCodeBench v5 and 70.8 on Berkeley Function Calling Leaderboard v3, with a CodeForces rating of 2,056 (developer evaluations, not interchangeable with non-thinking scores). Official cards warn that greedy thinking-mode decoding can degrade performance and cause endless repetitions, excessive repetition penalties can cause language mixing, and static YaRN can hurt shorter-text performance. Native 32K, configured 40,960 and YaRN-extended 128K are distinct; the launch blog lists 4B at 32K while its contemporaneous model card explicitly validates 128K with YaRN. Excludes later 2507 Instruct/Thinking refreshes, Qwen3-Coder, Next, VL and Omni. April 28 is provisional availability-by in UTC, supported by the official public-weight release README published that evening; the publisher labels the launch April 29, consistent with China time, and exact first public unlock remains unverified.",
      "paper_url": "https://arxiv.org/abs/2505.09388",
      "links": [
        "https://qwenlm.github.io/blog/qwen3/",
        "https://github.com/QwenLM/Qwen3",
        "https://www.alibabacloud.com/en/press-room/alibaba-introduces-qwen3-setting-new-benchmark?_p_lc=1",
        "https://huggingface.co/Qwen/Qwen3-0.6B",
        "https://huggingface.co/Qwen/Qwen3-0.6B-Base",
        "https://huggingface.co/Qwen/Qwen3-1.7B",
        "https://huggingface.co/Qwen/Qwen3-1.7B-Base",
        "https://huggingface.co/Qwen/Qwen3-4B",
        "https://huggingface.co/Qwen/Qwen3-4B-Base",
        "https://huggingface.co/Qwen/Qwen3-8B",
        "https://huggingface.co/Qwen/Qwen3-8B-Base",
        "https://huggingface.co/Qwen/Qwen3-14B",
        "https://huggingface.co/Qwen/Qwen3-14B-Base",
        "https://huggingface.co/Qwen/Qwen3-32B",
        "https://huggingface.co/Qwen/Qwen3-30B-A3B",
        "https://huggingface.co/Qwen/Qwen3-30B-A3B-Base",
        "https://huggingface.co/Qwen/Qwen3-235B-A22B",
        "https://modelscope.cn/models/Qwen/Qwen3-0.6B",
        "https://modelscope.cn/models/Qwen/Qwen3-0.6B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-1.7B",
        "https://modelscope.cn/models/Qwen/Qwen3-1.7B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-4B",
        "https://modelscope.cn/models/Qwen/Qwen3-4B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-8B",
        "https://modelscope.cn/models/Qwen/Qwen3-8B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-14B",
        "https://modelscope.cn/models/Qwen/Qwen3-14B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-32B",
        "https://modelscope.cn/models/Qwen/Qwen3-30B-A3B",
        "https://modelscope.cn/models/Qwen/Qwen3-30B-A3B-Base",
        "https://modelscope.cn/models/Qwen/Qwen3-235B-A22B"
      ],
      "page_url": "https://nlp.com.ai/models/qwen3-original-april-2025-release/",
      "record_updated_at": "2026-10-11T08:37:45.281779+00:00"
    },
    {
      "slug": "hyperclova-x-seed",
      "source_file": "2025-04-24-hyperclova-x-seed.json",
      "release_date": "2025-04-24",
      "date": "April 24, 2025",
      "name": "HyperCLOVA X SEED",
      "developer": "NAVER / NAVER Cloud (South Korea)",
      "parameters": "Text-Instruct-0.5B: 0.57B (566,281,216); Text-Instruct-1.5B: 1.59B (1,585,547,264); Vision-Instruct-3B: 3.72B total (3.26B text + 0.428B vision + 0.034B connector)",
      "context_window": "Text-Instruct-0.5B: 4K; Text-Instruct-1.5B and Vision-Instruct-3B: 16K",
      "license": "HyperCLOVA X SEED Model License Agreement (custom, not Apache 2.0); conditional commercial use, prohibited-use flow-down, license/NOTICE and modification notices, Powered by HyperCLOVA X attribution and derivative-model naming requirements; separate NAVER permission required above 10 million monthly active users at release or for competing products/services.",
      "key_features": "Original three-checkpoint instruction-tuned release: HyperCLOVAX-SEED-Text-Instruct-0.5B, HyperCLOVAX-SEED-Text-Instruct-1.5B, and HyperCLOVAX-SEED-Vision-Instruct-3B; benchmark-only Text-Base models and later Think/Omni releases are not included. NAVER-developed dense Transformer text generators focused on Korean language and culture; the 1.5B model supports Korean-English/Japanese translation, structured responses and business chat. The 0.5B model uses pruning and knowledge distillation followed by rejection-sampling and supervised fine-tuning; NAVER reports 4,358 A100 GPU hours and USD 6,537 for one pretraining run, excluding SFT, approximately 39 times lower than its Qwen2.5-0.5B comparison. The 1.5B model uses two-stage pretraining, rejection-sampling fine-tuning and SFT, with explicit long-context reinforcement. The 3B variant is a LLaVA-based vision-language generator combining a dense text decoder, SigLIP vision encoder and AnyRes C-Abstractor connector; it accepts text, images and video and generates text, with 378x378-pixel grids and up to nine grids. Its text post-training combines SFT and GRPO-based RLHF, with HyperCLOVA X-assisted data validation and vision-specific feedback. Publisher benchmarks show strong Korean performance and efficient video token use, but not universal superiority: the vision model trails Qwen2.5-VL-3B on VideoMME-Ko and PerceptionTest, and OCR/Lens auxiliary inputs improve image understanding. Published supported contexts are 4K/16K/16K, not the larger positional ceilings; the 0.5B config allows 8,192 positions and the vision text decoder 131,072, while its decoder_max_length is 16,384. The 1.5B weights require Hugging Face automatic access approval/contact sharing. April 24 is the official downloadable launch date, but the April 23-dated CLOVA download invitation conflicts with it; exact first public unlock remains unresolved. The linked later THINK report documents the original 0.5B pruning/distillation technique, not a shared THINK training recipe.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-0.5B",
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B",
        "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
        "https://www.navercorp.com/en/media/pressReleasesDetail?seq=32495",
        "https://www.navercorp.com/en/media/pressReleasesDetail?seq=32535",
        "https://clova.ai/en/tech-blog/sowing-the-seeds-in-the-ai-ecosystem-introducing-hyperclova-x-seed-a-commercial-open-source-ai",
        "https://clova.ai/cdn/media/2025/06/HyperCLOVA_X_THINK_Technical_Report.pdf"
      ],
      "page_url": "https://nlp.com.ai/models/hyperclova-x-seed/",
      "record_updated_at": "2026-10-11T08:37:45.279546+00:00"
    },
    {
      "slug": "kimi-vl",
      "source_file": "2025-04-09-kimi-vl.json",
      "release_date": "2025-04-09",
      "date": "April 9, 2025",
      "name": "Kimi-VL",
      "developer": "Moonshot AI (China)",
      "parameters": "16,407,657,776 total per checkpoint; approximately 16B total / 2.8B active in the MoE language decoder, plus a 400M MoonViT vision encoder and MLP projector",
      "context_window": "131,072 tokens (128K; both Instruct and original Thinking)",
      "license": "MIT (both original checkpoints; retain copyright and permission notices)",
      "key_features": "Original release groups Kimi-VL-A3B-Instruct and Kimi-VL-A3B-Thinking, not the later Thinking-2506 update. Generates text from text and visual inputs, including single/multiple images, documents and sampled video frames; it is not an image/video generator. Combines a Moonlight-derived MoE decoder with DeepSeek-style multi-head latent attention (MLA), 64 routed experts with six selected per token and two shared experts, a native-resolution SigLIP-derived MoonViT encoder using image packing and 2D RoPE, and a two-layer MLP projector with 2x2 spatial compression. Training uses an intermediate Moonlight checkpoint after 5.2T text tokens, then 4.4T tokens across vision training and joint text/multimodal pre-training, cooldown and long-context activation with an enhanced Muon optimizer; instruction tuning mixes text and visual dialogue, and Thinking adds long-CoT supervised fine-tuning and reinforcement learning with a response-length penalty. Moonshot recommends Instruct for efficient general perception, OCR, long documents/video and screenshot-based agents, and original Thinking for advanced text/visual reasoning. Publisher results for Instruct include 83.2 InfoVQA, 34.5 ScreenSpot-Pro, 64.5 LongVideoBench and 35.1 MMLongBench-Doc; original Thinking scores 61.7 MMMU, 36.8 MathVision and 71.3 MathVista-mini. These are task-specific strengths, not universal frontier parity: Instruct's OSWorld pass@1 is only 8.22%, and reported text/video needle recall drops to 87.0%/91.7% in the 65,536–131,072-token range. The original Thinking checkpoint is now deprecated in favor of the separately released June 2025 update.",
      "paper_url": "https://arxiv.org/abs/2504.07491",
      "links": [
        "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct",
        "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking",
        "https://modelscope.cn/models/moonshotai/Kimi-VL-A3B-Instruct",
        "https://modelscope.cn/models/moonshotai/Kimi-VL-A3B-Thinking",
        "https://github.com/MoonshotAI/Kimi-VL",
        "https://x.com/Kimi_Moonshot/status/1910035354570371082"
      ],
      "page_url": "https://nlp.com.ai/models/kimi-vl/",
      "record_updated_at": "2026-10-11T08:37:45.277552+00:00"
    },
    {
      "slug": "llama-31-nemotron-ultra-253b-v1",
      "source_file": "2025-04-07-llama-3-1-nemotron-ultra-253b-v1.json",
      "release_date": "2025-04-07",
      "date": "April 7, 2025",
      "name": "Llama-3.1-Nemotron-Ultra-253B-v1",
      "developer": "NVIDIA (USA)",
      "parameters": "253B (253,401,268,224 parameters; dense)",
      "context_window": "128K (131,072 tokens)",
      "license": "NVIDIA Open Model License + underlying Llama 3.1 Community License Agreement and Acceptable Use Policy; commercial use subject to terms (not Apache 2.0)",
      "key_features": "Text-only reasoning and chat model derived from Meta Llama-3.1-405B-Instruct, not Llama 4. Puzzle neural architecture search compresses the dense decoder with skipped attention, variable-width FFNs and FFN fusion; recovery uses 65B knowledge-distillation tokens followed by 88B continual-pretraining tokens. Supervised fine-tuning covers math, code, reasoning, chat and tool calling, followed by multiple GRPO reinforcement-learning stages and alignment. Reasoning is toggled with the system prompt detailed thinking on/off; NVIDIA recommends placing other instructions in the user prompt. Intended for agents, RAG, tool use and English/code, with additional supported languages. NVIDIA reports reasoning-on scores of 76.01 GPQA, 72.50 AIME25, 66.31 LiveCodeBench (2024-08-01 to 2025-02-01), 97.00 MATH500 and 74.10 BFCL V2 Live, evaluated at 32K sequence length; reasoning-off scores are substantially lower on math, science and coding. The original downloadable checkpoint is BF16 and NVIDIA lists 8x H100-80GB or 4x B100 test hardware; memory, KV-cache and serving settings limit practical concurrency/context. The separate CPT base checkpoint and later official FP8 packaging are not grouped into this launch card.",
      "paper_url": "https://arxiv.org/abs/2505.00949",
      "links": [
        "https://huggingface.co/nvidia/Llama-3_1-Nemotron-Ultra-253B-v1",
        "https://developer.nvidia.com/blog/build-enterprise-ai-agents-with-advanced-open-nvidia-llama-nemotron-reasoning-models/",
        "https://github.com/NVIDIA/TensorRT-LLM",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE"
      ],
      "page_url": "https://nlp.com.ai/models/llama-31-nemotron-ultra-253b-v1/",
      "record_updated_at": "2026-10-11T08:37:45.275866+00:00"
    },
    {
      "slug": "llama-4",
      "source_file": "2025-04-05-llama-4.json",
      "release_date": "2025-04-05",
      "date": "April 5, 2025",
      "name": "Llama 4",
      "developer": "Meta",
      "parameters": "Scout: 109B total / 17B active, 16 routed experts; Maverick: 400B nominal total / 17B active, 128 routed experts (MoE; full BF16 checkpoints approximately 108.64B and 401.58B parameters respectively)",
      "context_window": "Base Scout and Maverick: 256K; Scout Instruct: 10M; Maverick Instruct: 1M tokens",
      "license": "Llama 4 Community License Agreement + Acceptable Use Policy (custom commercial open-weight terms, not Apache/MIT; attribution and derivative naming obligations; separate Meta permission above 700M monthly active users; multimodal rights excluded for EU-domiciled individuals and EU-principal-place-of-business companies, with an incorporated-product end-user exception)",
      "key_features": "Same-day launch family comprising Scout and Maverick base and instruction-tuned checkpoints, with official Maverick Instruct FP8 packaging; unreleased Behemoth is excluded. Autoregressive, natively multimodal early-fusion models accept multilingual text and images and generate text/code; video-frame training is not a separate native video or audio interface. Scout uses MoE in every decoder layer, while Maverick alternates dense and MoE layers; each MoE token activates one routed expert plus a shared expert, but all experts must be stored. The full-checkpoint parameter totals include the multimodal components, not just a 17B dense language model. iRoPE interleaves chunked RoPE attention with global no-position-embedding attention and uses inference-time attention temperature scaling for long-context generalization. Meta reports approximately 40T pretraining tokens for Scout and 22T for Maverick, 200 pretraining languages with 12 explicitly supported, and an August 2024 knowledge cutoff. Long-context mid-training and a lightweight SFT, online RL, lightweight DPO post-training recipe improve reasoning and multimodal chat; Maverick was co-distilled from the Behemoth teacher. Base context is 256K; the 10M/1M Instruct limits are not universal training lengths or guarantees of reliable task performance at every position. Meta reports BF16 Instruct MMMU scores of 69.4/73.4 and GPQA Diamond scores of 57.2/69.8 for Scout/Maverick; these are developer evaluations, not independent verification. Scout prioritizes long-document processing and image grounding, while Maverick reports stronger visual reasoning and coding results. Scout fits a single H100 only with on-the-fly INT4 quantization; Maverick FP8 fits an H100 DGX host, not one H100 GPU, and long contexts need additional cache memory. The launch LMArena Elo 1417 belongs to an experimental chat variant, not the downloadable Maverick checkpoint. Outputs may be inaccurate, and deployment requires application-specific safety testing and guardrails.",
      "paper_url": "",
      "links": [
        "https://ai.meta.com/blog/llama-4-multimodal-intelligence/",
        "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E",
        "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct",
        "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E",
        "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct",
        "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md",
        "https://huggingface.co/blog/llama4-release",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama4/USE_POLICY.md"
      ],
      "page_url": "https://nlp.com.ai/models/llama-4/",
      "record_updated_at": "2026-10-11T08:37:45.273876+00:00"
    },
    {
      "slug": "deepseek-v3-0324",
      "source_file": "2025-03-24-deepseek-v3-0324.json",
      "release_date": "2025-03-24",
      "date": "March 24, 2025",
      "name": "DeepSeek-V3-0324",
      "developer": "DeepSeek AI (China)",
      "parameters": "671B total / 37B active per token (MoE; auxiliary MTP weights additional)",
      "context_window": "128K supported; 163,840 tokens configured",
      "license": "MIT",
      "key_features": "Updated text-only V3 chat checkpoint with the same DeepSeekMoE and Multi-head Latent Attention (MLA) architecture: 61 main layers, 256 routed experts with 8 selected per token and one shared expert, FP8 weights, and one auxiliary multi-token prediction (MTP) module. The cited V3 report describes the original backbone trained on 14.8T tokens with an auxiliary-loss-free load-balancing strategy, MTP objective, supervised fine-tuning and reinforcement learning; it does not document the March update's training recipe. DeepSeek reports improvements over V3 on MMLU-Pro (75.9 to 81.2), GPQA (59.1 to 68.4), AIME (39.6 to 59.4) and LiveCodeBench (39.2 to 49.2), alongside more executable and visually polished web/game front-ends, stronger Chinese long-form writing, rewriting, translation and search-result analysis, and more accurate function calling. Supports JSON output and fill-in-the-middle completion; search and uploaded-file workflows use supplied text prompts rather than native browsing or multimodal input. DeepSeek recommends V3 for non-complex reasoning tasks. Published 128K V3 support and the checkpoint's 163,840-position YaRN configuration are distinct; performance across the full configured limit is not established. The release card cautions that direct Hugging Face Transformers support was not yet available and directs users to dedicated local inference frameworks. Weights and repository are MIT-licensed.",
      "paper_url": "https://arxiv.org/abs/2412.19437",
      "links": [
        "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324",
        "https://github.com/deepseek-ai/DeepSeek-V3",
        "https://api-docs.deepseek.com/news/news250325/"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-v3-0324/",
      "record_updated_at": "2026-10-11T08:37:45.271650+00:00"
    },
    {
      "slug": "llama-nemotron-nano-8b-and-super-49b-v1",
      "source_file": "2025-03-18-llama-nemotron.json",
      "release_date": "2025-03-18",
      "date": "March 18, 2025",
      "name": "Llama Nemotron (Nano 8B and Super 49B v1)",
      "developer": "NVIDIA (USA)",
      "parameters": "Nano: 8B (8,030,261,248); Super: 49B (49,867,145,216); dense, not MoE",
      "context_window": "131,072 tokens (128K), both variants",
      "license": "NVIDIA Open Model License, with underlying Llama 3.1 Community License (Nano) / Llama 3.3 Community License (Super) and their Acceptable Use Policies; commercial use subject to attribution, redistribution, naming and >700M-MAU additional-license conditions (not Apache 2.0)",
      "key_features": "Text-only reasoning and chat family comprising Llama-3.1-Nemotron-Nano-8B-v1 and Llama-3.3-Nemotron-Super-49B-v1. Nano is a dense decoder-only Transformer post-trained from Llama 3.1 8B Instruct for local RTX/edge use. Super compresses Llama 3.3 70B Instruct using distillation-based neural architecture search, with 31 of 80 blocks skipping attention and variable-width feed-forward blocks; it remains dense, not a Mamba hybrid or mixture of experts. Super received 40B knowledge-distillation tokens from FineWeb, Buzz-V1.2 and Dolma. Both use supervised fine-tuning for math, code, reasoning and tool calling, followed by RLOO and online reward-aware preference optimization, with final SFT/RPO checkpoint merging. A system prompt switches detailed thinking on or off for reasoning versus lower-overhead chat, RAG and tool use. NVIDIA reports reasoning-on MATH500 scores of 95.4% for Nano and 96.6% for Super, and AIME25 scores of 47.1% and 58.4%, respectively, evaluated at 32K sequence length. Reasoning is task-dependent: Nano strict-instruction IFEval falls from 82.1% with reasoning off to 79.3% on, and thinking can still occur when disabled. Primarily intended for English and code, with seven additional supported languages; base pretraining knowledge cutoff is 2023. Super targets an accuracy/throughput tradeoff, but the released BF16 checkpoint lists two 80GB GPUs for inference rather than guaranteeing single-GPU deployment. Ultra is not part of this launch-day weight release.",
      "paper_url": "https://arxiv.org/abs/2505.00949",
      "links": [
        "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-8B-v1",
        "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1",
        "https://nvidianews.nvidia.com/news/nvidia-launches-family-of-open-reasoning-ai-models-for-developers-and-enterprises-to-build-agentic-ai-platforms",
        "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/",
        "https://arxiv.org/abs/2505.00949"
      ],
      "page_url": "https://nlp.com.ai/models/llama-nemotron-nano-8b-and-super-49b-v1/",
      "record_updated_at": "2026-10-11T08:37:45.269225+00:00"
    },
    {
      "slug": "mistral-small-31",
      "source_file": "2025-03-17-mistral-small-3-1.json",
      "release_date": "2025-03-17",
      "date": "March 17, 2025",
      "name": "Mistral Small 3.1",
      "developer": "Mistral AI (France)",
      "parameters": "24B each (Base and Instruct; dense multimodal; approximately 23.57B text model + 403M vision encoder + 35.65M connector)",
      "context_window": "128K (131,072 tokens)",
      "license": "Apache 2.0",
      "key_features": "March 2025 multimodal release with pretrained Base and instruction-finetuned Instruct checkpoints, adding image understanding and extending context to 128K while improving text performance over Mistral Small 3. Dense autoregressive architecture combines a 40-layer Mistral text decoder with grouped-query attention (32 query heads, 8 KV heads), a 24-layer Pixtral vision encoder and a patch-merging connector; uses the 131,072-token Tekken vocabulary. Supports text and image inputs with text output, dozens of languages, long-document understanding, coding and mathematical reasoning; Instruct adds native function calling, JSON output and system-prompt support. Mistral reports 88.41% HumanEval, 94.08% DocVQA and 81.20% RULER at 128K for Instruct, with strong multimodal and multilingual performance in its size class. Performance is not uniformly best: reported MATH is 69.30% versus Gemma 3 27B IT's 89.00%, and RULER falls from 93.96% at 32K to 81.20% at 128K. Mistral advertises inference speeds of 150 tokens/s; single-RTX-4090 or 32GB-Mac deployment requires quantization, whereas the model card estimates roughly 55GB GPU RAM for BF16/FP16. Base is pretrained-only, intended for completion and downstream fine-tuning rather than ready-made instruction following. Training data composition, token budget and detailed post-training recipe are not disclosed in the checked release sources.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/mistral-small-3-1/",
        "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Base-2503",
        "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-small-31/",
      "record_updated_at": "2026-10-11T08:37:45.266397+00:00"
    },
    {
      "slug": "command-a",
      "source_file": "2025-03-13-command-a.json",
      "release_date": "2025-03-13",
      "date": "March 13, 2025",
      "name": "Command A",
      "developer": "Cohere and Cohere Labs (Canada)",
      "parameters": "111B dense (111,057,580,032 checkpoint parameters)",
      "context_window": "256K supported; 128K configured by default on Hugging Face",
      "license": "CC-BY-NC-4.0 with Cohere Labs Acceptable Use Addendum (noncommercial; attribution required)",
      "key_features": "Text-only, instruction-tuned enterprise language model with a dense decoder-only Transformer architecture, grouped-query attention, SwiGLU, shared input/output embeddings, and a repeating 3:1 pattern of 4,096-token sliding-window attention with RoPE and full attention without positional embeddings. Pretrained on multilingual web text and code, synthetic datasets, human-annotated instruction data, and vendor-sourced data; post-training combines supervised fine-tuning, domain-specialized expert weight merging, preference optimization, reinforcement learning, and final conversational polishing into a single checkpoint, not a mixture-of-experts inference model. Trained for 23 languages, citation-grounded RAG, multistep tool use and agents, SQL generation, code translation, and long-document understanding, with contextual and strict safety modes. Cohere reports performance competitive with GPT-4o and DeepSeek-V3 on enterprise-focused human evaluations, 150% higher throughput than Command R+ 08-2024, and deployment on two A100/H100 GPUs; these are publisher claims, not guarantees for every serving configuration. The technical report records MATH 80.0, Bird-SQL 59.5, and Taubench 51.7, but BFCL 63.8 trails GPT-4o's 72.1, and RULER performance falls from 97.2 at 4K to 84.6 at 256K. The released checkpoint supports 256K context, although its Hugging Face configuration defaults to 128K and must be changed for longer inputs; the hosted API's 8K output limit is separate. Default responses are conversational and verbose, so extraction or terse-answer workflows may need explicit system instructions. Open weights are a gated, noncommercial research release with attribution and acceptable-use requirements.",
      "paper_url": "https://arxiv.org/abs/2504.00698",
      "links": [
        "https://huggingface.co/CohereLabs/c4ai-command-a-03-2025",
        "https://cohere.com/blog/command-a",
        "https://docs.cohere.com/docs/command-a",
        "https://cohere.com/cohere-labs-cc-by-nc-license",
        "https://docs.cohere.com/docs/cohere-labs-acceptable-use-policy"
      ],
      "page_url": "https://nlp.com.ai/models/command-a/",
      "record_updated_at": "2026-10-11T08:37:45.263593+00:00"
    },
    {
      "slug": "gemma-3",
      "source_file": "2025-03-12-gemma-3.json",
      "release_date": "2025-03-12",
      "date": "March 12, 2025",
      "name": "Gemma 3",
      "developer": "Google DeepMind",
      "parameters": "1B / 4B / 12B / 27B (dense language-model sizes; 4B, 12B and 27B additionally use a shared 417M-parameter SigLIP vision encoder)",
      "context_window": "32K tokens (1B); 128K tokens (4B, 12B and 27B)",
      "license": "Gemma Terms of Use (custom open-weight terms with use restrictions and redistribution obligations; not a permissive Apache/MIT license)",
      "key_features": "Launch family of four sizes, each with pre-trained (PT) and instruction-tuned (IT) checkpoints. The 1B model is text-only; 4B, 12B and 27B accept interleaved text and images and generate text. Dense decoder-only Transformers use grouped-query attention, RMSNorm and QK normalization, with five local sliding-window attention layers per global layer and a 1,024-token local window to reduce long-context KV-cache memory. The larger models share a frozen SigLIP vision encoder operating at 896 x 896 resolution and compressing each image or crop into 256 soft tokens; inference-time Pan & Scan improves high-resolution and non-square image understanding. Knowledge-distillation pre-training uses 2T / 4T / 12T / 14T tokens for 1B / 4B / 12B / 27B, with multilingual data covering over 140 languages. Instruction tuning combines distillation, reinforcement learning from human, machine and execution feedback, and model merging to improve math, coding, chat and instruction following; Google also highlights structured outputs and function calling. Supports question answering, summarization, visual data extraction and multilingual tasks. The report describes 4B-IT as competitive with Gemma 2 27B-IT and 27B-IT as comparable to Gemini 1.5 Pro across benchmarks, and reports a preliminary text-only Chatbot Arena Elo of 1338 for 27B-IT. Limitations include incorrect or outdated factual statements, training-data biases, difficulty with highly complex or ambiguous tasks, and safety evaluations limited to English prompts; the report observes rapid degradation when context is extended beyond 128K. ShieldGemma 2 is a separate image-safety classifier, not an included Gemma 3 checkpoint.",
      "paper_url": "https://arxiv.org/abs/2503.19786",
      "links": [
        "https://developers.googleblog.com/en/introducing-gemma3/",
        "https://ai.google.dev/gemma/docs/core/model_card_3",
        "https://huggingface.co/google/gemma-3-1b-pt",
        "https://huggingface.co/google/gemma-3-1b-it",
        "https://huggingface.co/google/gemma-3-4b-pt",
        "https://huggingface.co/google/gemma-3-4b-it",
        "https://huggingface.co/google/gemma-3-12b-pt",
        "https://huggingface.co/google/gemma-3-12b-it",
        "https://huggingface.co/google/gemma-3-27b-pt",
        "https://huggingface.co/google/gemma-3-27b-it",
        "https://ai.google.dev/gemma/terms",
        "https://arxiv.org/abs/2503.19786"
      ],
      "page_url": "https://nlp.com.ai/models/gemma-3/",
      "record_updated_at": "2026-10-11T08:37:45.261015+00:00"
    },
    {
      "slug": "aya-vision",
      "source_file": "2025-03-04-aya-vision.json",
      "release_date": "2025-03-04",
      "date": "March 4, 2025",
      "name": "Aya Vision",
      "developer": "Cohere Labs (Canada)",
      "parameters": "8B and 32B (dense; approximately 8.63B and 33.14B total checkpoint parameters including vision components)",
      "context_window": "16K (both variants)",
      "license": "CC BY-NC 4.0 with Cohere Labs Acceptable Use Addendum (both variants)",
      "key_features": "Open-weight multilingual vision-language research family with 8B and 32B variants, accepting text and images and generating text across 23 languages. Supports OCR, image captioning, visual question answering, chart and document understanding, visual reasoning, summarization, translation, and screenshot-to-code tasks, alongside text-only chat. The 8B language backbone starts from Command R7B and is post-trained with the Aya Expanse recipe; the 32B backbone starts from Aya Expanse 32B. Both use a SigLIP2 vision encoder and a two-layer SwiGLU multimodal connector with Pixel Shuffle to reduce image-token counts fourfold; the paper specifies patch14-384 for 8B and patch16-512 for 32B, with dynamic image tiling up to 12 tiles plus a thumbnail. Training combines vision-language alignment, multilingual supervised fine-tuning, synthetic re-annotation, translation and rephrasing, and cross-modal model merging to preserve text-only capabilities. Released with AyaVisionBench and multilingual m-WildVision evaluations; the authors report strong multilingual open-ended preference performance, including 32B competing with substantially larger vision-language models. These preference gains do not imply universal benchmark superiority: the paper reports lower xMMMU accuracy than the compared Qwen2.5-VL models (39.94 versus 46.77 for 8B/7B; 45.11 versus 61.74 for 32B/72B). Both checkpoints have a published 16K context and are restricted to non-commercial use under an acceptable-use addendum; Hugging Face downloads require accepting access conditions.",
      "paper_url": "https://arxiv.org/abs/2505.08751",
      "links": [
        "https://huggingface.co/CohereLabs/aya-vision-8b",
        "https://huggingface.co/CohereLabs/aya-vision-32b",
        "https://cohere.com/blog/aya-vision",
        "https://huggingface.co/blog/aya-vision",
        "https://cohere.com/cohere-labs-cc-by-nc-license"
      ],
      "page_url": "https://nlp.com.ai/models/aya-vision/",
      "record_updated_at": "2026-10-11T08:37:45.258292+00:00"
    },
    {
      "slug": "phi-4-mini-and-phi-4-multimodal",
      "source_file": "2025-02-26-phi-4-mini-and-multimodal.json",
      "release_date": "2025-02-26",
      "date": "February 26, 2025",
      "name": "Phi-4-mini and Phi-4-multimodal",
      "developer": "Microsoft Research (United States)",
      "parameters": "3.84B (Phi-4-mini, dense); 5.57B (Phi-4-multimodal, Mixture-of-LoRAs)",
      "context_window": "131,072 tokens",
      "license": "MIT",
      "key_features": "Same-day release of the compact Phi-4-mini-instruct text model and Phi-4-multimodal-instruct. Both use a 32-layer language backbone with a 3,072-dimensional hidden state, 24 attention heads, eight key-value heads, and a 200,064-token vocabulary. Phi-4-mini is a dense decoder-only Transformer with grouped-query attention and shared input-output embeddings, designed for multilingual reasoning, mathematics, coding, instruction following, and function calling in compute-constrained or latency-sensitive environments. Its model card lists support for 24 text languages and describes post-training with supervised fine-tuning and direct preference optimization. Phi-4-multimodal extends the language backbone through a Mixture-of-LoRAs architecture that unifies text, image, and audio inputs while generating text output. It supports simultaneous audio-and-vision reasoning, image and document understanding, OCR, charts and tables, multiple-image comparison, speech recognition, speech translation, speech question answering, speech summarization, and general audio understanding. Text supports the same 24-language set; vision is documented for English, while audio supports English, Chinese, German, French, Italian, Japanese, Spanish, and Portuguese. Multimodal post-training used supervised fine-tuning, direct preference optimization, and reinforcement learning from human feedback. At launch Microsoft reported a 6.14% word-error rate on the Hugging Face OpenASR leaderboard, performance above WhisperV3 and SeamlessM4T-v2-Large on its reported speech-recognition and translation evaluations, and competitive document, chart, OCR, mathematical, and scientific visual reasoning despite the 5.6B model size. Both models target on-device and edge deployment and are released with standard Transformers checkpoints as well as Microsoft-supported deployment paths.",
      "paper_url": "https://arxiv.org/abs/2503.01743",
      "links": [
        "https://azure.microsoft.com/en-us/blog/empowering-innovation-the-next-generation-of-the-phi-family/",
        "https://huggingface.co/microsoft/Phi-4-mini-instruct",
        "https://huggingface.co/microsoft/Phi-4-multimodal-instruct",
        "https://modelscope.cn/models/microsoft/Phi-4-mini-instruct",
        "https://modelscope.cn/models/microsoft/Phi-4-multimodal-instruct",
        "https://github.com/microsoft/PhiCookBook"
      ],
      "page_url": "https://nlp.com.ai/models/phi-4-mini-and-phi-4-multimodal/",
      "record_updated_at": "2026-10-11T08:37:45.255442+00:00"
    },
    {
      "slug": "wayfarer-large-70b-llama-33",
      "source_file": "2025-02-19-wayfarer-large-70b-llama-3-3.json",
      "release_date": "2025-02-19",
      "date": "February 19, 2025",
      "name": "Wayfarer-Large-70B-Llama-3.3",
      "developer": "Latitude (USA), in collaboration with Gryphe Padar",
      "parameters": "70B dense (70,553,706,496 parameters)",
      "context_window": "Unknown (131,072 tokens configured; finetune-supported maximum unverified)",
      "license": "Llama 3.3 Community License Agreement + incorporated Acceptable Use Policy (custom license, not Apache 2.0; redistribution/attribution/naming requirements and >700 million MAU commercial condition)",
      "key_features": "Text-only adventure-roleplay finetune of Meta Llama 3.3 70B Instruct, with 80 decoder layers, 8,192 hidden width and grouped-query attention using 64 query heads and 8 key/value heads. Latitude describes an approximately equal 33/33/33 mixture of 8K text-adventure data, 4K roleplay data and a SlimOrca Sonnet instruction subset; the latter helped distinguish instructions from fiction and amplify the intended negative sentiment. Adventure data came from simulated AI Dungeon character-creator scenarios, with five user archetypes and separate narrator/player language models; simulations ran up to 8K tokens or until the character died. Data was regenerated and improved after the original 12B release to reduce repetition and improve narrative flow. The publisher emphasizes challenging combat, failure, death and consequences rather than positivity bias or plot armor. Exact optimization stages and any DPO/RL step are not disclosed for this checkpoint. Training is specialized for second-person present-tense narration; official guidance warns that third-person consistency, gentler stories and acting for the player can be difficult, with repetition or stalled pacing requiring user guidance. The 131,072-position configuration is not a published long-context validation; training sample lengths and hosted subscription context limits are distinct. Uses the original Llama 3 header-based chat template, not ChatML, with end-of-turn token <|eot_id|>; recommended local sampling baseline is temperature 0.8, min_p 0.05 and repetition_penalty 1.05. Public BF16 safetensors and publisher-endorsed GGUF downloads allow local use without an AI Dungeon subscription; the subscription requirement concerns hosted Wayfarer Large only. No quantitative checkpoint benchmarks or dedicated paper were found. February 19 is a provisional public-release recommendation, not a verified exact first weight-access date; contemporary local execution is documented by February 20.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Wayfarer-Large-70B-Llama-3.3",
        "https://huggingface.co/LatitudeGames/Wayfarer-Large-70B-Llama-3.3-GGUF",
        "https://help.aidungeon.com/ai-model-differences",
        "https://huggingface.co/LatitudeGames/Wayfarer-Large-70B-Llama-3.3/discussions/2",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE",
        "https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/USE_POLICY.md"
      ],
      "page_url": "https://nlp.com.ai/models/wayfarer-large-70b-llama-33/",
      "record_updated_at": "2026-10-11T08:37:45.252603+00:00"
    },
    {
      "slug": "mistral-small-3",
      "source_file": "2025-01-30-mistral-small-3.json",
      "release_date": "2025-01-30",
      "date": "January 30, 2025",
      "name": "Mistral Small 3",
      "developer": "Mistral AI (France)",
      "parameters": "23.6B (dense; marketed as 24B)",
      "context_window": "32,768 tokens",
      "license": "Apache 2.0",
      "key_features": "Latency-optimized dense language-model release containing both Mistral-Small-24B-Base-2501 and Mistral-Small-24B-Instruct-2501. The architecture has 40 Transformer layers, a 5,120-dimensional hidden state, 32 attention heads, eight key-value heads, and a 131,072-token Tekken vocabulary. The instruction checkpoint supports multilingual conversation, system prompts, native function calling, and JSON output; listed languages include English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Russian, and Korean, while the model card additionally identifies Dutch and Polish support. Mistral positioned the model for fast conversational agents, low-latency tool use, domain-specific fine-tuning, and private local inference. A quantized checkpoint can fit on a single RTX 4090 or a Mac with 32GB of memory, while the BF16 instruction checkpoint requires approximately 55GB of GPU memory. Mistral reported 80.73 on MMLU for the base checkpoint and instruction-model results of 66.3 on MMLU-Pro, 45.3 on GPQA, 84.8 on HumanEval, 70.6 on MATH, and 8.35 on MT-Bench. The launch announcement reported more than 81% MMLU accuracy and approximately 150 tokens per second on an RTX 4090, describing the model as more than three times faster than Llama 3.3 70B on the same hardware. Mistral states that Small 3 was trained without reinforcement learning or synthetic data, leaving the base relatively early in the model-production pipeline for downstream customization.",
      "paper_url": "",
      "links": [
        "https://mistral.ai/news/mistral-small-3/",
        "https://huggingface.co/mistralai/Mistral-Small-24B-Base-2501",
        "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501",
        "https://modelscope.cn/models/mistralai/Mistral-Small-24B-Base-2501",
        "https://modelscope.cn/models/mistralai/Mistral-Small-24B-Instruct-2501"
      ],
      "page_url": "https://nlp.com.ai/models/mistral-small-3/",
      "record_updated_at": "2026-10-11T08:37:45.249905+00:00"
    },
    {
      "slug": "deepseek-r1",
      "source_file": "2025-01-20-deepseek-r1.json",
      "release_date": "2025-01-20",
      "date": "January 20, 2025",
      "name": "DeepSeek-R1",
      "developer": "DeepSeek (China)",
      "parameters": "671B total / 37B active (R1 and R1-Zero MoE); 1.5B, 7B, 8B, 14B, 32B, and 70B (dense distills)",
      "context_window": "128K tokens",
      "license": "MIT (distilled checkpoints also retain applicable Qwen or Llama base-model terms)",
      "key_features": "First-generation DeepSeek reasoning-model family built from DeepSeek-V3-Base. DeepSeek-R1-Zero applied large-scale reinforcement learning directly to the base model without preliminary supervised fine-tuning, producing emergent self-verification, reflection, and long-chain-of-thought behavior but also repetition, readability, and language-mixing problems. DeepSeek-R1 addresses those weaknesses by introducing cold-start data before reinforcement learning and using a pipeline with two reinforcement-learning stages and two supervised-fine-tuning stages for reasoning, preference alignment, and non-reasoning behavior. Both full models use a 671-billion-parameter mixture-of-experts architecture with 37 billion parameters activated per token. DeepSeek also released six dense distilled checkpoints based on Qwen2.5 and Llama 3 models, trained on reasoning samples generated by R1. Official R1 results include 90.8 on MMLU, 84.0 on MMLU-Pro, 71.5 on GPQA Diamond, 65.9 on LiveCodeBench, a 2,029 Codeforces rating, 49.2% resolved on SWE-bench Verified, 79.8 on AIME 2024, and 97.3 on MATH-500. The official evaluation used sampling for applicable benchmarks with temperature 0.6, top-p 0.95, 64 responses per problem for pass-at-one estimation, and a maximum generation length of 32,768 tokens.",
      "paper_url": "https://arxiv.org/abs/2501.12948",
      "links": [
        "https://github.com/deepseek-ai/DeepSeek-R1",
        "https://huggingface.co/deepseek-ai/DeepSeek-R1",
        "https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero",
        "https://modelscope.cn/models/deepseek-ai/DeepSeek-R1"
      ],
      "page_url": "https://nlp.com.ai/models/deepseek-r1/",
      "record_updated_at": "2026-10-11T08:37:45.247108+00:00"
    },
    {
      "slug": "wayfarer-12b",
      "source_file": "2025-01-16-wayfarer-12b.json",
      "release_date": "2025-01-16",
      "date": "January 16, 2025",
      "name": "Wayfarer-12B",
      "developer": "LatitudeGames (USA)",
      "parameters": "12.248B (12,247,802,880; dense)",
      "context_window": "131,072 tokens (configured; finetune support unverified)",
      "license": "Apache 2.0",
      "key_features": "English adventure-storytelling and roleplay finetune of Mistral-Nemo-Base-2407, developed by Latitude with Gryphe Padar; dense 40-layer Mistral causal Transformer with grouped-query attention and BF16 weights. Designed for AI Dungeon adventures with conflict, dangerous choices, frequent failure and character death rather than guaranteed success or plot armor. Uses two-stage supervised fine-tuning in ChatML: 180K chat-formatted instruction examples, followed by a 50/50 mixture of synthetic 8K-context text adventures and roleplay experiences. Adventure data simulates published AI Dungeon character-creator scenarios with five user archetypes and separate narrator/user language models blind to each other’s logic; simulations run up to 8K tokens or character death. The model card reports exclusively second-person present-tense narrative and single-turn chat training, with other narrative styles potentially suboptimal; Latitude also acknowledges underuse of the end-of-sequence token, so explicit generation-length limits matter. Recommended baseline sampling is temperature 0.8, repetition penalty 1.05 and min_p 0.025. Official GGUF conversions are available. The checkpoint config permits 131,072 positions, but no Wayfarer-specific validated maximum is published; 8K training simulations and hosted service limits do not establish its maximum context. This is the original January 2025 model, not Wayfarer-2, Wayfarer Large or Muse.",
      "paper_url": "",
      "links": [
        "https://huggingface.co/LatitudeGames/Wayfarer-12B",
        "https://huggingface.co/LatitudeGames/Wayfarer-12B-GGUF",
        "https://www.reddit.com/r/LocalLLaMA/comments/1i2t82i/introducing_wayfarer_a_brutally_challenging/",
        "https://huggingface.co/LatitudeGames/Wayfarer-12B/discussions/4"
      ],
      "page_url": "https://nlp.com.ai/models/wayfarer-12b/",
      "record_updated_at": "2026-10-11T08:37:45.244281+00:00"
    },
    {
      "slug": "minimax-01",
      "source_file": "2025-01-15-minimax-01.json",
      "release_date": "2025-01-15",
      "date": "January 15, 2025",
      "name": "MiniMax-01",
      "developer": "MiniMax (China)",
      "parameters": "456B total / 45.9B active (MoE; Text-01 and VL-01; VL-01 adds a 303M ViT)",
      "context_window": "Up to 4,000,000 tokens (Text-01 inference; trained at 1,000,000)",
      "license": "MiniMax Model License (custom)",
      "key_features": "Foundation-model family comprising MiniMax-Text-01 and MiniMax-VL-01. MiniMax-Text-01 is an 80-layer mixture-of-experts model with 32 experts and top-2 routing, a 6,144-dimensional hidden state, 64 attention heads, and a 200,064-token vocabulary. Its hybrid architecture places one conventional softmax-attention layer after every seven Lightning Attention layers, combining linear-complexity long-context processing with periodic full attention. MiniMax used Linear Attention Sequence Parallelism Plus, variable-length ring attention, Expert Tensor Parallelism, and computation-communication overlap to train at a one-million-token context and support inference at up to four million tokens. Official results include 88.5 on MMLU, 75.7 on MMLU-Pro, 89.1 on IFEval, 56.5 on LongBench v2 with chain of thought, and a 0.910 RULER score at one million tokens. MiniMax-VL-01 adds a 24-layer, 303-million-parameter Vision Transformer and a two-layer MLP projector to the Text-01 backbone. It supports dynamic image resolutions from 336 by 336 through 2016 by 2016, retains a thumbnail view, and processes tiled patches for detailed image understanding. Its vision encoder was trained from scratch on 694 million image-caption pairs, while the four-stage multimodal pipeline processed 512 billion vision-language tokens. Reported VL scores include 68.5 on MMMU, 91.7 on ChartQA, 96.4 on DocVQA, and 865 on OCRBench.",
      "paper_url": "https://arxiv.org/abs/2501.08313",
      "links": [
        "https://github.com/MiniMax-AI/MiniMax-01",
        "https://huggingface.co/MiniMaxAI/MiniMax-Text-01",
        "https://huggingface.co/MiniMaxAI/MiniMax-VL-01",
        "https://huggingface.co/MiniMaxAI/MiniMax-Text-01/blob/main/LICENSE-MODEL"
      ],
      "page_url": "https://nlp.com.ai/models/minimax-01/",
      "record_updated_at": "2026-10-11T08:37:45.241511+00:00"
    },
    {
      "slug": "sky-t1-32b-preview",
      "source_file": "2025-01-10-sky-t1-32b-preview.json",
      "release_date": "2025-01-10",
      "date": "January 10, 2025",
      "name": "Sky-T1-32B-Preview",
      "developer": "NovaSky / UC Berkeley Sky Computing Lab (United States)",
      "parameters": "32.8B (dense; Qwen2.5-32B-Instruct fine-tune)",
      "context_window": "32,768 tokens",
      "license": "Apache 2.0",
      "key_features": "Open-weight reasoning model produced by supervised fine-tuning Qwen2.5-32B-Instruct on a released dataset of approximately 17,000 verified reasoning traces. The training mixture covers mathematics, competitive programming, science, and puzzles using material derived from APPs, TACO, NuminaMath, AIME, MATH, Olympiad problems, and Still-2. NovaSky generated candidate traces with QwQ-32B-Preview, rejected mathematically incorrect answers and code that failed provided unit tests, and used GPT-4o-mini to rewrite traces into a more consistently parseable format. Fine-tuning ran for three epochs with a batch size of 96 and completed in 19 hours on eight H100 GPUs using LlamaFactory and DeepSpeed ZeRO-3 Offload; the team estimated the compute cost at less than $450 using Lambda Cloud pricing. Reported scores include 82.4 on MATH-500, 43.3 on AIME 2024, 56.8 on GPQA Diamond, and 86.3, 56.8, and 17.9 on the easy, medium, and hard LiveCodeBench partitions. The release includes model weights, its training dataset, data-curation recipes, training code, evaluation tooling, and experiment logs.",
      "paper_url": "",
      "links": [
        "https://novasky-ai.github.io/posts/sky-t1/",
        "https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview",
        "https://huggingface.co/datasets/NovaSky-AI/Sky-T1_data_17k",
        "https://github.com/NovaSky-AI/SkyThought",
        "https://modelscope.cn/models/NovaSky-AI/Sky-T1-32B-Preview"
      ],
      "page_url": "https://nlp.com.ai/models/sky-t1-32b-preview/",
      "record_updated_at": "2026-10-11T08:37:45.238884+00:00"
    },
    {
      "slug": "phi-4",
      "source_file": "2024-12-12-phi-4.json",
      "release_date": "2024-12-12",
      "date": "December 12, 2024",
      "name": "Phi-4",
      "developer": "Microsoft Research (United States)",
      "parameters": "14.7B (dense; marketed as 14B)",
      "context_window": "16,384 tokens",
      "license": "MIT",
      "key_features": "Open-weight, primarily English reasoning model built as a 40-layer dense decoder-only Transformer with a 5,120-dimensional hidden state, 40 attention heads, 10 key-value heads, and a 100,352-token vocabulary. Microsoft trained Phi-4 on 9.8 trillion tokens over 21 days using 1,920 H100-80GB GPUs. Its data mixture emphasizes synthetic textbook-style material for mathematics, coding, common-sense reasoning, and general knowledge, supplemented by filtered public web documents, selected educational and code data, acquired academic books and question-answer datasets, and supervised chat data. Post-training combined supervised fine-tuning with iterative direct preference optimization. In Microsoft's SimpleEval results, Phi-4 scored 84.8 on MMLU, 56.1 on GPQA, 80.4 on MATH, and 82.6 on HumanEval; its GPQA and MATH scores exceeded the reported Llama 3.3 70B Instruct results, although its 3.0 SimpleQA score exposed substantially weaker factual recall. The model is intended for reasoning-intensive, latency-sensitive, and memory-constrained text applications.",
      "paper_url": "https://arxiv.org/abs/2412.08905",
      "links": [
        "https://huggingface.co/microsoft/phi-4",
        "https://modelscope.cn/models/microsoft/phi-4"
      ],
      "page_url": "https://nlp.com.ai/models/phi-4/",
      "record_updated_at": "2026-10-11T08:37:45.235378+00:00"
    }
  ]
}