Open Weight Text-to-Text Models

Browse a chronological table of open-weight large language model (LLM) releases, with parameter counts, licenses, technical summaries, research papers, model weights, and other primary sources. Download the catalogue as JSONL, JSON, or CSV.

Download dataset: JSONL JSON CSV Schema Agent guide

Tracked releases by month

Line plot of tracked open-weight model releases by month. The current partial month is marked separately.
Date Model Developer Parameters License Key Features Paper
November 2026 11 releases Month to date
November 15, 2026 Agnes 2.5 Pro Theta Agnes AI (Singapore) 187.76B total / approximately 7.81B active for text generation (approximately 8.26B including the vision encoder) Apache License 2.0 Open-weight multimodal mixture-of-experts checkpoint for conversational reasoning, scientific problem solving, coding, long-context work, and visual understanding. Agnes 2.5 Pro Theta accepts text, images, and video frames and generates text, with 187,763,545,616 BF16 parameters and 52 registered text decoder layers. Inspection of the released tensor headers and execution code gives approximately …
November 12, 2026 Inkling 2 Small (NVFP4, MXFP8) Thinking Machines Lab, Inc. (United States) 594B total / 17B active (sparse MoE; 8 of 384 experts per token) Apache 2.0 Open-weight general-purpose multimodal model accepting text and image inputs and generating text, with particular emphasis on agentic coding, tool use, knowledge work, visual reasoning, long-context operation, retrieval-augmented generation, and conversational applications. Its 48-layer decoder-only transformer uses a sparse Mixture-of-Experts feed-forward backbone with each token routed to eight of 384 experts, …
November 11, 2026 Raon-1 nano KRAFTON AI (South Korea) 3.1B main / 1.5B effective (dense; plus a 55M MTP module) Apache 2.0 Open-weight Korean- and English-capable instruction model designed for natural conversation, long-context instruction following, reasoning, and tool calling on memory-constrained and on-device hardware. Raon-1 nano is a dense model with 3.1 billion logical parameters but approximately 1.5 billion effective parameters; roughly 1.5B of its parameters reside in per-layer token-embedding tables. It … Link
November 10, 2026 Nex-N2.6 Nex AGI (China) 321.3B total / approximately 18B active (sparse MoE; FP8 checkpoint) Apache 2.0 Open-weight multimodal agentic model post-trained from GLM-5.3-Flash for computer use, coding, long-horizon workflows, and tool-driven tasks. Nex-N2.6 retains the GLM-5.3-Flash architecture with approximately 321.3 billion total parameters and an approximately 18-billion-parameter active path. Its 45-layer, 4,096-dimensional language backbone contains three initial dense feed-forward layers followed by 42 sparse Mixture-of-Experts layers. …
November 10, 2026 Bielik-11B-v3.1-Instruct SpeakLeash Foundation and ACK Cyfronet AGH (Poland) 11.17B (11,168,796,672; dense) Apache 2.0 Open-weight Polish-first multilingual instruction model expanding the Bielik 11B v3 family from 32 to 56 supported languages, including every official language of the European Union as well as Arabic, Chinese, Japanese, Korean, Armenian, Azerbaijani, Georgian, Hebrew, Hindi, Indonesian, Persian, Vietnamese, and other regional languages. Bielik-11B-v3.1-Instruct is a dense causal decoder …
November 10, 2026 ARLoop The Ohio State University and Saarland University (USA/Germany) 0.6B and 1.2B physical parameters; three recurrent passes producing 54 and 72 effective layers Apache 2.0 Fully open family of pretrained recurrent-depth language models introducing Attention Residuals for looped Transformers. Rather than repeatedly accumulating every intermediate representation into a single residual stream, each effective sublayer uses a learned query to selectively attend over representations produced at earlier depths. Residual updates are grouped into blocks to control … Link
November 9, 2026 LongCat-2.5 Meituan LongCat (China) 1.6T total / approximately 48B active per token (sparse MoE) MIT Open-weight native multimodal Mixture-of-Experts model designed for long-horizon agentic coding, full-project development, creative design, 3D work, visual programming, research, and general tool use. LongCat-2.5 contains 1.6 trillion total parameters and activates approximately 48 billion parameters per token. The released configuration uses a 38-layer, 8,192-dimensional language backbone with Multi-head Latent Attention, …
November 6, 2026 Arete-27B Nous Research (USA) 27B dense (Qwen3.8-27B fine-tune) Apache 2.0 Nous Research post-trained Alibaba's Qwen3.8-27B to alter refusals, deflections, and Chinese-state framing that it observed on politically sensitive topics, describing the objective as reinforcing Western values while demonstrating that open-weight models can be adapted to different value systems. Arete retains the dense Qwen3.8-27B hybrid architecture and native 262,144-token context window, …
November 5, 2026 ERNIE-5.5-Flash-Base Baidu ERNIE Team (China) 280B total / 14B active per token, plus 0.36B ECHO parameters ERNIE 5.5 Flash Base Model … Open-weight, text-only sparse mixture-of-experts base model released at the end of pretraining, before any mid-training, instruction tuning, reinforcement learning, or behavioral post-training. ERNIE-5.5-Flash-Base was pretrained from scratch on 33.5 trillion tokens with the Muon optimizer and is intended as a foundation for domain adaptation, instruction tuning, and reinforcement-learning research rather …
November 4, 2026 Ripple-65B Preview Ricursive Intelligence (USA) 65B (dense) Apache License 2.0 Open-weight, text-only general-purpose preview derived from the proprietary model powering Ricursive Intelligence's AI chip-design platform. Ripple-65B Preview is a dense Transformer pretrained from scratch on 12 trillion predominantly English and code tokens, followed by 200B mid-training tokens, dedicated long-context training, 183B supervised-fine-tuning tokens, long-context supervised fine tuning, and mixed-domain reinforcement …
November 4, 2026 Kimi K3.1 Pro Moonshot AI (China) 2.8T total / 104B active (MoE) Kimi K3.1-Pro License Open-weight, natively multimodal agentic flagship and first release in Moonshot AI's K3.1 series, post-trained from Kimi K3 using substantially greater post-training compute and large-scale joint reinforcement learning across hundreds of downstream domains. The 93-layer Mixture-of-Experts architecture combines 69 Kimi Delta Attention layers with 24 Gated DeepSeek Sparse Attention layers, replacing …
October 2026 36 releases
October 30, 2026 SenseNova-Loop-9B-Preview SenseNova / SenseTime (China) 9.66B stored / approximately 14.6B inference parameters (dense shared-weight looped model; exact checkpoint inventory: 9,656,250,480 parameters) Apache-2.0 Preview open-weight multimodal language model for long-context document understanding, coding, tool use, and agentic workflows. It accepts text and images and generates text, with optional thinking controlled through its chat template. Instead of increasing stored parameter count conventionally, SenseNova Loop reuses a 24-layer decoder block twice: layers 0-3 execute once, …
October 29, 2026 Mistral Large 4 (Le Chonk) Mistral AI (France) 1.05T total / 52B active (49B active transformer parameters per token plus embeddings and output layers; 1.6B vision encoder) Mistral Ecosystem Open-Weig… Final open-weight release of Mistral Large 4 following its October 6 public preview and three additional weeks of reinforcement learning. It is a natively multimodal text-and-image-input, text-output model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. The 72-layer granular mixture-of-experts language backbone has 128 routed …
October 27, 2026 Spark-X2.6-1.7B SparkLLM / XHToken (Singapore) 1.708B dense (1,707,657,216) Apache-2.0 Second model in the Spark-X2.6 series and a distinct later release from Spark-X2.6-4B. The compact English/Chinese text model retains exactly the same architecture as Spark-X2.5-1.7B and attributes its capability gains to improved post-training rather than architectural scaling. Its 28-layer dense Transformer repeats three 512-token sliding-attention layers followed by one full-attention …
October 27, 2026 Ling-3.1-flash-1019 inclusionAI (China) 560B total / 25B active (572.1B serialized; BF16 base and blockwise-FP8 post-trained checkpoints) MIT Coordinated open-weight release comprising the Ling-3.1-flash-base pretrained checkpoint and the separately post-trained Ling-3.1-flash-1019 model. The base checkpoint supports continued pretraining, domain adaptation, supervised post-training, reinforcement learning, and research; the 1019 checkpoint adds further undisclosed post-training aimed at reasoning and real-world agentic tasks. Both use the same core hybrid-linear Mixture-of-Experts architecture …
October 27, 2026 LFM3 Liquid AI (USA) 1.73B text/base; 1.92B vision and d1; 2B audio; 2.17B omni and d1-Omni LFM Open License v1.0 (comm… Same-day open-weight family of compact text, vision, audio, omni-modal, and non-generative decision models built on Liquid AI's new tri-hybrid backbone. The post-trained LFM3-1.7B text checkpoint contains exactly 1,728,801,536 BF16 parameters in a single approximately 3.46 GB safetensors file and uses 24 layers comprising 13 gated short-convolution blocks, five 256-token sliding-attention …
October 25, 2026 Nanbeige4.3-33B Nanbeige LLM Lab / BOSS Zhipin (China) 33.119B (33,118,977,024; dense) Apache 2.0 (model-card decl… Open-weight dense text-generation model positioned as the flagship teacher for the Nanbeige4.3 family and the foundation for distilling agentic capabilities into forthcoming smaller models such as Nanbeige4.3-3B. The BF16 checkpoint uses a Llama-compatible causal architecture with 52 transformer layers, a 7,168-dimensional hidden state, 22,016-dimensional feed-forward layers, 56 query heads, eight …
October 23, 2026 NVIDIA Nemotron 3.5 Bolt E4B NVIDIA Corporation (USA) 8B total / 4B effective (dense Gemma 4 E4B) OpenMDW License Agreement, … Open-weight compact text-generation model further post-trained from Gemma 4 E4B IT for fast local conversational agents, voice and device assistants, interactive characters, and robotic, industrial, and embedded interfaces. Despite its E4B designation, the checkpoint contains approximately 8.0B parameters; NVIDIA describes it as 4B effective and markets approximately 4GB as a … Link
October 22, 2026 MiniMax-M3.1-Flash MiniMax (China) ~428B total / ~23B active (MoE) minimax-community Open-weight native multimodal Mixture-of-Experts model for coding, long-horizon agents, computer use, cybersecurity, and professional workflows, supporting text, image, and video input with text output. MiniMax-M3.1-Flash retains M3's approximately 428B-total / 23B-active scale while introducing MiniMax Sparse Attention throughout all 60 transformer layers, replacing the three full-attention layers retained by M3. … Link
October 22, 2026 Ling-3.0-flash-Sante inclusionAI (China) 124B total / 5.1B active (MoE) MIT MIT-licensed healthcare-specialized version of Ling-3.0-flash designed for medical reasoning, healthcare research, evidence synthesis, deep research, tool-assisted workflows, and multi-agent collaboration. It retains the BailingMoeV3 hybrid-linear Mixture-of-Experts architecture with 42 transformer layers, including 35 Kimi Delta Attention layers and seven gated Multi-head Latent Attention layers, two initial dense layers, 512 routed …
October 22, 2026 Carbon-X Hugging Face Bio (United States) / Zhongguancun Academy (China) / TIGEM and University of Naples Federico II (Italy) 14.8B (dense; 14,822,046,720 parameters) Apache-2.0 Open-weight biological language model derived from Qwen3-14B and specialized for joint reasoning over natural language and raw DNA. Carbon-X retains Qwen3's dense 40-layer transformer backbone while introducing a HybridDNATokenizer that combines the original Qwen vocabulary with 5,515 DNA-specific tokens, including all 4,096 six-base combinations, shorter tail tokens, DNA boundary markers, …
October 21, 2026 Flint Recursive Superintelligence (United States) 39.8B total / approximately 2.5B active Apache-2.0 for the model we… Apache-2.0 decoder-only Mixture-of-Experts language model designed for local coding agents that investigate repositories, fix bugs, implement features, write tests, reason, and use tools over multiple turns. Flint was designed and trained from scratch by Recursive Superintelligence in under four months with assistance from the company's Eureka Machine automated research system. …
October 21, 2026 Raon-1 mini preview KRAFTON AI (South Korea) 28.8B total / 4.3B active Apache-2.0 for downloadable… Preview release of a Korean- and English-capable reasoning and agentic language model designed and trained from scratch by KRAFTON AI. Raon-1 mini is a sparse Mixture-of-Experts model with 28.8B total parameters and 4.3B activated per token, described by KRAFTON as a 30B-A4B-class model. Its 25-layer hybrid architecture contains 18 fixed-state …
October 20, 2026 MemT-1.0 MemTensor (China) 36B total / 3B active (Nano); 288B total / 17B active (Flash) Apache-2.0 (Nano); MIT (Fla… Family of memory-native text-generation models that makes persistent, dynamically evolving memory part of model computation rather than repeatedly supplying the original conversation history at inference. The release includes MemT-1.0-Nano, built on Qwen3.6-35B-A3B with 36B total and 3B activated parameters, and MemT-1.0-Flash, built on a further-post-trained DeepSeek-V4-Flash backbone with 288B total … Link
October 20, 2026 Spark-X2.6-4B SparkLLM / XHToken (Singapore) 4.11B dense Apache-2.0 Apache-2.0 compact general-purpose language model for local and on-device agents. Spark-X2.6-4B retains exactly the same 4.11B-parameter architecture as Spark-X2.5-4B and attributes its improvements to a new post-training recipe rather than additional parameters or architectural changes. The 36-layer dense Transformer uses a 2,560-dimensional hidden state, 10,240-dimensional feed-forward layers, 16 query heads, …
October 20, 2026 Xing4.0-29B-Intent China Telecom Artificial Intelligence Technology Co., Ltd. (China) 29B total / 4B active (MoE) Apache-2.0 Apache-2.0 task-specialized fine-tune of Xing4.0-29B-A4B for schema-guided intent recognition and structured information extraction. Xing4.0-29B-Intent uses a Unified Schema-driven Semantic Parsing protocol that represents intent recognition and slot filling, named-entity recognition, relation extraction, and event extraction through the same category-plus-args JSON structure. Users provide candidate categories, descriptions, and extractable field definitions; …
October 20, 2026 Ornith-1.5-122B Ornith Team 122B MoE (approximately 125B including multimodal components) MIT MIT-licensed addition to the Ornith-1.5 family, released two months after the original 9B, 35B-A3B, and 397B checkpoints. Ornith-1.5-122B is a native image-and-video-capable reasoning and agent model built with the Qwen3.5 MoE conditional-generation architecture. Its configuration contains 48 language-model layers with a 3,072-dimensional hidden state, 32 query heads, 2 key-value heads, …
October 19, 2026 Ming-3.0-audio inclusionAI / Ant Group (China) 126B total / 6.5B active (Flash); 9.3B total / 2.7B active (Tiny), both MoE MIT Open-weight native audio-language family comprising Ming-3.0-audio-flash and Ming-3.0-audio-tiny. Both accept audio together with text instructions and produce text or optional synthesized speech, supporting automatic speech recognition, spoken question answering, long-recording analysis, non-speech sound and acoustic-scene understanding, music understanding, audio reasoning, spoken interaction, and audio-conditioned tool calling. Flash uses a 42-layer …
October 17, 2026 DeepSeek-V4.1-Flash-1017 DeepSeek AI (China) 552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE) MIT 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 … Link
October 16, 2026 NVIDIA Nemotron Labs Dory 1.0 4B NVIDIA Corporation (USA) 3.98B physical parameters (looped dense model) OpenMDW License Agreement, … 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 …
October 16, 2026 Beam 1.0 Reflection AI 501.3B total / 23.4B active (sparse MoE) Apache 2.0 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 … Link
October 15, 2026 Hy-Memory-30B-A3B Tencent Hunyuan (China) 30B total, approximately 3B active Apache License 2.0 Open-weight mixture-of-experts model purpose-trained as a long-term-memory writer for LLM applications. Hy-Memory-30B-A3B replaces the general-purpose language model used along the write path of Mem0, MemOS, and Hy-Memory while leaving each framework's schemas, retrieval logic, reader model, and memory-bank implementation intact. It decides what information should persist, preserves exact names, numbers, … Link
October 15, 2026 Ternary Bonsai 2 2B / 4B / 9B Prism ML, Inc. (United States) 1.88B / 4.21B / 8.95B language model + 0.33B / 0.33B / 0.46B vision tower Apache 2.0 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 … Link
October 15, 2026 Muse Spark Open Meta Superintelligence Labs (USA) 816B total / 42B active MoE (approximately 822.0B serialized BF16 parameters) Muse Model License 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 … Link
October 15, 2026 Comet 60B Multiverse Computing (Spain) 61.1B total / 12.3B active (marketed as 60B total / 12B active MoE) NVIDIA Nemotron Open Model … 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, …
October 15, 2026 NVIDIA Nemotron 3.5 Bolt 74B-A9B NVIDIA Corporation (USA) 74B total / 9B active (LatentMoE) OpenMDW License Agreement, … 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 … Link
October 15, 2026 GPT-USENET-6 HDTenEightyP 0.4B dense (421,772,288 serialized parameters) MIT License 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 …
October 15, 2026 NVIDIA Nemotron 3.5 Super VL 120B-A12B NVIDIA Corporation (USA) 120B total / 12B active (multimodal MoE) OpenMDW License Agreement, … 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 … Link
October 15, 2026 Jais 3 Institute of Foundation Models, MBZUAI / Inception42 (UAE) 0.9B, 3.7B, 7B, and 32B dense; 36B total / 4B active (MoE + MoVA); 375B total / 23B active (MoE) Apache License 2.0 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 …
October 15, 2026 Step 5 Preview StepFun (China) 600B total / 27B active (MoE) StepFun Model License 1.0 (… 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, …
October 14, 2026 BeeLLM 2.0 schizophyllume (skizy) 1.1M and 11,965,760 weights (dense Llama architecture) Apache 2.0 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 …
October 13, 2026 Ling-3.1-flash inclusionAI (China) 560B total / 25B active (hybrid-linear MoE) MIT 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 …
October 13, 2026 LFM2.5-24B-A2B Liquid AI (USA) 24.0B total / 2.45B active (MoE) LFM Open License v1.0 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 … Link
October 12, 2026 MiniCPM-V 4.7 (35B-A3B, 1B) OpenBMB / ModelBest (China) 35B-A3B: 35.21B stored / approximately 3B active (MoE); 1B: 1.55B stored (1.3B-class dense) Apache 2.0 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 …
October 12, 2026 ColGrep-Agent-2B LightOn (France) 2.52B dense (2,516,756,480 BF16 parameters; marketed as 2B) Apache 2.0 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 … Link
October 6, 2026 Mr.Chatterbox-2.7B-Chat Trip Venturella 2.7B dense (2,705,820,160 serialized parameters) CC BY-NC-SA 4.0 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 …
October 3, 2026 Kolibri-1 Aleph Alpha (Germany) 78.1B total / 3.46B active (MoE) Apache 2.0 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 … Link
September 2026 19 releases
September 27, 2026 Naive-N0.5-Flash NaiveAI 309B total / 15.5B active (MoE) MIT 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, …
September 24, 2026 Krake V2 Legacy NovelAI / Anlatan (United States) 21B (dense; GPT-NeoX-20B fine-tune) GPL-3.0 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 …
September 24, 2026 Clio V1 Legacy NovelAI / Anlatan (United States) 3B (dense) GPL-2.0-only 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 …
September 22, 2026 MiMo-V2.6 (MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, MiMo-V2.6-Distill-Qwen-9B) Xiaomi MiMo (China) 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 MIT 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 … Link
September 21, 2026 Tinfield 1 Bad Theory Labs (Nigeria) 177B total / 6.6B active per token (Sparse MoE; approximately 180B stored parameters) Qwen Community License 1.0 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 …
September 20, 2026 Hemmingway-1 Altworld (Switzerland / South Africa) 27B dense (26,895,998,464 BF16 parameters) Apache 2.0 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 …
September 18, 2026 AliceAI-Foundation-80B-A3B-Base Yandex (Russia) 80B total / 3B active (MoE; 81,286,433,408 stored BF16 parameters) Apache 2.0 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 …
September 17, 2026 Xing4.0-29B-A4B China Telecom Artificial Intelligence Technology Co., Ltd. (China) 29B total / 4B active (MoE) Apache-2.0 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 …
September 17, 2026 Ternary Bonsai 2 27B Prism ML, Inc. (United States) 27.36B total: 24.35B language backbone + 2.54B embeddings/LM head + 0.46B vision tower Apache 2.0 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, … Link
September 15, 2026 Zero (MC001) Moving Castles 8B (dense) Not specified 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 … Link
September 14, 2026 Atria Dawn Preview (BF16, FP8) Shanghai Artificial Intelligence Laboratory (China) 753B checkpoint / ~40B active (MoE; post-trained from the 744B-A40B GLM-5.2 foundation) MIT 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 … Link
September 11, 2026 Agnes-3.0-Flash Agnes AI (Singapore) 33.09B (dense) Apache License 2.0 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 …
September 10, 2026 DeepSeek-V4.1-Flash DeepSeek AI (China) 552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE) MIT 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. … Link
September 8, 2026 Nex-N2.5 (Max, mini) Nex AGI (China) 1.6T total (Max); 35B total (mini) Apache-2.0 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 …
September 8, 2026 Ling-3.0-flash-VL inclusionAI (China) 124B total / 5.5B active (MoE) MIT 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, …
September 7, 2026 MiniCPM5-2B (Final, SFT, Midtrain, Base, GGUF, MLX, GPTQ, DSpark) OpenBMB / ModelBest (China) 2,516,756,480 Apache 2.0 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 … Link
September 3, 2026 Ling-3.0-flash-Fin inclusionAI (China) 124B total / 5.1B active (MoE; 127.49B stored parameters) MIT 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 …
September 3, 2026 K2 Horizon (0.9B, 3.7B, 7B, 32B Stage1, 36B-A4B, 375B-A23B) Institute of Foundation Models / MBZUAI (United Arab Emirates) 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 Apache 2.0 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 …
September 1, 2026 Spark-X2.5 (1.7B and 4B) SparkLLM / XHToken (Singapore) 1.71B and 4.11B (dense; marketed as 1.7B and 4B) Apache 2.0 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 …
August 2026 26 releases
August 31, 2026 DeepSeek-V4-Flash-Vision-Exp DeepSeek AI (China) 305B total (MoE) MIT 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 …
August 28, 2026 Hy4 preview (BF16 + FP8) Tencent Hy Team (China) 770B total / 49B active MoE backbone + 10B total / 0.7B active MTP layer Apache 2.0 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 …
August 28, 2026 GLM-5.3 (FP8) Z.AI / Zhipu AI (China) ~753B total / 40B active (MoE; same base as GLM-5.2, with the official GLM family specification also described as 744B-A40B) GLM-5.3 License 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 … Link
August 26, 2026 Qwen3.8-Flash-Next Alibaba Qwen Team (China) 125B total / 6B active MoE backbone, with 51B N-gram embeddings and 4B MTP parameters Qwen Community License 1.0 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 … Link
August 26, 2026 GLM-5.3-Flash (FP8, BF16) Z.AI / Zhipu AI (China) 320B total / 18B active (MoE; official checkpoint contains approximately 321.3B parameters) MIT 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 … Link
August 25, 2026 Granite 4.2 (3B, 8B, 30B) IBM (USA) 3.66B, 8.79B, 29.28B dense (marketed as 3B, 8B, 30B) Apache License 2.0 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 …
August 24, 2026 Thomson-1.0-Small Thomson Reuters (Canada) 35.107B total / approximately 3B active (MoE) PolyForm Strict License 1.0… 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 … Link
August 19, 2026 Ornith-1.5 Ornith Team 9B dense; 35B total / approximately 3B active MoE; 397B MoE MIT 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 …
August 19, 2026 Ling-3.0 Base Checkpoints (Tiny and Flash) inclusionAI (China) 7.9B total / 1.3B active (Tiny); 124B total / 5.1B active (Flash) MIT 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 … Link
August 14, 2026 Qwen3.8-27B Alibaba Qwen Team (China) 27B dense language model; 27,781,427,952 total BF16 parameters including vision encoder Apache License 2.0 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 …
August 14, 2026 Intern-S2-Mobius (BF16, FP8) Shanghai AI Laboratory / InternLM (China) 35B-class sparse multimodal model; 35,964,405,616 parameters in the BF16 checkpoint Apache-2.0 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 … Link
August 13, 2026 dots3-note Preview dots studio / Xiaohongshu (China) 280B total / 16B active (multimodal MoE) Apache License 2.0 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 …
August 13, 2026 DeepSeek-V4-Pro-0813 DeepSeek AI (China) 1.65T checkpoint / 49B active target model (MoE; includes attached DSpark module) MIT 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 … Link
August 12, 2026 Qwen3.8-2.4T-A95B Alibaba Qwen Team (China) 2.4T total / 95B active (MoE) Qwen3.8-Max License 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 …
August 12, 2026 North Micro Vision Instruct Cohere / Cohere Labs (Canada) 2.4B total (2B language model + 400M vision encoder) Apache 2.0 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 …
August 12, 2026 LFM2.5-VL-3B Liquid AI (USA) 3.1B total (LFM2.5-2.6B language backbone + 400M SigLIP2 NaFlex vision encoder) LFM Open License v1.0 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 … Link
August 11, 2026 NVIDIA Nemotron 3.5 Lightning 30B-A3B NVIDIA Corporation (USA) 30B total / 3B active (MoE) OpenMDW License Agreement, … 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; …
August 10, 2026 Muse Glimmer 30B Meta Superintelligence Labs (USA) Approximately 29.6B including approximately 1.8B perception encoder Apache License 2.0 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 … Link
August 10, 2026 Motif-3 Motif Technologies (South Korea) ~314B total / ~13.2B active per token (sparse MoE) MIT License 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 … Link
August 10, 2026 Ling-3.0-tiny inclusionAI (China) 7.9B total / 1.3B active (MoE) MIT 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. …
August 5, 2026 Mach-1 Additive 35B (Text, Multimodal / Mach-1 Small) Syzygy Research 35B total / approximately 3B active (MoE) Apache License 2.0 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 …
August 4, 2026 Maple-Preview / Maple-Preview (Flash) DeepGrove (USA) 20.2B total / 1.49B active (MoE) MIT 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 …
August 4, 2026 Ling-3.0-flash inclusionAI (China) 124B total / 5.1B active (MoE) MIT 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 …
August 4, 2026 LFM2.5-2.6B (Base, Agentic) Liquid AI (USA) 2.69B LFM Open License v1.0 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. … Link
August 2, 2026 BigBang-V1 The Endless Frontier 35.95B total / approximately 3B active (MoE) Apache License 2.0 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 …
August 1, 2026 LongCat-Flash-Lite-Sparse Meituan LongCat (China) 69B total / ~3B active per token (MoE) MIT 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 … Link
July 2026 26 releases
July 31, 2026 openPangu-2.0-Pro Huawei (China) 505B total / 18B active (MoE) OpenPangu Model License Agr… 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, … Link
July 31, 2026 DeepSeek-V4-Flash-0731 DeepSeek AI (China) 284B total / 13B active (MoE) MIT 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 … Link
July 31, 2026 BarunLM-35M Harshal Singh (India) 35,072,768 Apache License 2.0 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 …
July 30, 2026 Inkling-Small (BF16, NVFP4) Thinking Machines Lab, Inc. (United States) 276B total / 12B active (MoE) Apache 2.0 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 …
July 29, 2026 A.X K2 (FP8) SK Telecom (South Korea) 688B total / 33B active (MoE) Apache 2.0 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 … Link
July 27, 2026 Neutrino-1 (8B, 0.6B, 0.6B-Chat) Fermion Research 8.19B; 596M (0.6B and 0.6B-Chat) Apache 2.0 weights; bundled… 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, …
July 27, 2026 Kimi K3 Moonshot AI (China) 2.8T total, 104B active (MoE) Kimi K3 License 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 … Link
July 24, 2026 Instella-MoE-16B-A3B (Pretrain, Midtrain, Base, SFT, DPO, Think) AMD (USA) 16B total / 2.8B active per token (MoE) Research-Only RAIL Model Li… 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 …
July 23, 2026 KAT-Coder-V2.5-Dev (BF16) KwaiKAT Team at Kuaishou Technology (China) 35B total / 3B active (MoE; 34.66B checkpoint parameters) Apache 2.0 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 … Link
July 22, 2026 Solar Open 2 (250B-A15B) Upstage (South Korea) 250B total / 15B active per token (MoE) Upstage Solar License 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 … Link
July 22, 2026 Nanbeige4.2-3B Nanbeige (China) 4B total / 3B non-embedding parameters Apache-2.0 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 …
July 21, 2026 Macaron-V1 (Venti, Coding-Venti, Tall) Mind Lab Venti/Coding-Venti: 748B aggregate (744B base + four 1B LoRA specialists), ~40B active; Tall: 50B aggregate (35B-A3B base + four 3.7B LoRA specialists) MIT 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 …
July 21, 2026 Laguna S 2.1 (BF16, FP8, NVFP4, INT4, GGUF, DFlash) Poolside (United States) 118B total / 8B active per token (MoE) OpenMDW-1.1 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 …
July 21, 2026 Antares (350M, 1B) Cisco Foundation AI / Cisco Systems, Inc. (United States) 350M and 1B dense Apache 2.0 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 … Link
July 20, 2026 Motif-3-Beta Motif Technologies (South Korea) ~314B total / ~13B active per token (sparse MoE) Custom non-commercial resea… 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 …
July 20, 2026 BTL-3 Bad Theory Labs (Nigeria) 27B (dense) Apache 2.0 model artifacts;… 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 … Link
July 17, 2026 Intern-S2-Preview-397B (BF16, FP8) Shanghai AI Laboratory / InternLM (China) 397B-to-403B class multimodal model (Hugging Face lists approximately 403B parameters) Apache-2.0 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 … Link
July 16, 2026 LLaDA2.2-flash inclusionAI / Ant Group (China) 103B total (100B non-embedding, MoE) Apache 2.0 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 … Link
July 15, 2026 Inkling (BF16, NVFP4) Thinking Machines Lab, Inc. (United States) 975B total / 41B active (MoE) Apache 2.0 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, …
July 14, 2026 Bonsai 27B (1-bit, ternary, AWQ, unpacked) Prism ML, Inc. 27.3B language parameters + ~0.46B vision tower Apache 2.0 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 … Link
July 14, 2026 Agents-A1-4B (base, FP8, GGUF) InternScience 4B dense Apache 2.0 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 … Link
July 10, 2026 Soofi S 30B-A3B (Base, Instruct Preview, Isar Preview, Rhine Preview) Soofi Consortium (Germany) 30B total / ~3B active (hybrid MoE; Hugging Face cards also cite ~3.5B active) Custom / Other (paper says … 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 … Link
July 6, 2026 Nemotron-Labs-3-Puzzle-75B-A9B (BF16, FP8, NVFP4) NVIDIA (United States) 75.3B total / 9.3B active (hybrid MoE) OpenMDW-1.1 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 … Link
July 6, 2026 Hy3 (BF16 + FP8) Tencent Hy Team (China) 295B total / 21B active (MoE) + 3.8B MTP layer Apache 2.0 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, …
July 5, 2026 LongCat-2.0 Meituan LongCat (China) 1.6T total / ~48B active per token (MoE) MIT 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 …
July 2, 2026 Laguna XS 2.1 (BF16, FP8, NVFP4, INT4, GGUF) Poolside (United States) 33B total / 3B active per token (MoE) OpenMDW-1.1 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 …
June 2026 14 releases
June 25, 2026 Ornith-1.0 DeepReinforce 9B dense, 35B MoE, 397B MoE publicly confirmed; official release page also describes a 31B dense variant MIT 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 …
June 25, 2026 Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 NVIDIA (United States) ~63.2B total shipped parameters across two towers (about 30B AR/context tower + 30B diffusion/denoiser tower); ~3B active per tower NVIDIA Nemotron Open Model … 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 … Link
June 25, 2026 LFM2.5-230M Liquid AI (USA) 230M (LFM2 hybrid architecture; 14 layers: 8 double-gated LIV conv + 6 GQA) LFM Open License v1.0 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 … Link
June 24, 2026 Qwen-AgentWorld-35B-A3B Alibaba Qwen Team (China) 35B total / 3B active (MoE) Apache-2.0 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. … Link
June 16, 2026 GLM-5.2 (BF16, FP8) Z.AI / Zhipu AI (China) ~753B total / 40B active (MoE; official repo table also labels the family 744B-A40B) MIT 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, … Link
June 12, 2026 Rio 3.5 Open 397B IplanRIO / Prefeitura do Rio de Janeiro (Brazil) 397B total / 17B active (MoE-derived release) MIT 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 …
June 12, 2026 MiniMax-M3 MiniMax (China) ~428B total / ~23B active (MoE) minimax-community 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 … Link
June 12, 2026 Kimi-K2.7-Code Moonshot AI (China) 1T total / 32B active Modified MIT License 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 …
June 10, 2026 DiffusionGemma-26B-A4B-it Google DeepMind (USA) 25.2B total / 3.8B active (MoE, marketed as A4B) Apache-2.0 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 … Link
June 9, 2026 North Mini Code Cohere / Cohere Labs (Canada) 30B total / 3B active (MoE) Apache 2.0 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 …
June 8, 2026 Quasar-Preview SILX AI ~18B total / ~2B active (MoE) MIT 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 …
June 8, 2026 Nex-N2 (Pro, mini) Nex AGI (China) 397B / A17B (Pro); 35B / A3B (mini) Apache-2.0 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, …
June 8, 2026 Macaron-V1-Preview-749B Mind Lab 749B total (744B base + 5B LoRA adapters) / 40B active MIT 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, … Link
June 4, 2026 Nemotron 3 Ultra 550B-A55B (Base BF16, BF16, NVFP4) NVIDIA (USA) 550B total / 55B active (MoE) OpenMDW-1.1 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, … Link
May 2026 13 releases
May 29, 2026 Step 3.7 Flash StepFun (China) 198B total / ~11B active (196B language + 1.8B vision) Apache 2.0 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 …
May 28, 2026 LFM2.5-8B-A1B (Base, Reasoning-tuned) Liquid AI (USA) 8.3B total / 1.5B active (MoE) LFM 1.0 license 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 … Link
May 26, 2026 Mellum2-12B-A2.5B (Base, Instruct, Thinking) JetBrains 12B total / 2.5B active (MoE) Apache 2.0 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 … Link
May 26, 2026 Laguna M.1 Poolside (USA) 225.8B total / 23.4B active (MoE) Apache 2.0 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 … Link
May 21, 2026 Supra-50M (Base, Instruct) SupraLabs 50M Apache 2.0 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 …
May 21, 2026 MiniCPM5-1B (Final, SFT) OpenBMB 1,080,632,832 Apache 2.0 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, … Link
May 20, 2026 Command A+ Cohere (Canada) 25B active / 218B total (MoE) Apache 2.0 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 …
May 19, 2026 Nemotron-Labs-Diffusion (3B, 8B, 14B; Base + Instruct + VLM-8B) NVIDIA (United States) 3B / 8B / 14B dense LM family; VLM-8B is listed as 9B on Hugging Face NVIDIA Nemotron Open Model … 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 … Link
May 18, 2026 HRM-Text-1B Sapient Intelligence 1B Apache-2.0 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 … Link
May 18, 2026 Equinox-31B LatitudeGames 31B Apache 2.0 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 …
May 15, 2026 Ring-2.6-1T inclusionAI / Ant Group (China) 1T MIT 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 … Link
May 15, 2026 Intern-S2-Preview Shanghai AI Laboratory / InternLM (China) 35B-A3B-class MoE; 36B total parameters in Hugging Face metadata Apache-2.0 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, … Link
May 6, 2026 ZAYA1-8B (ZAYA1-8B + ZAYA1-reasoning-base) Zyphra (USA) 8.4B total / 760M active (MoE) Apache 2.0 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 … Link
April 2026 19 releases
April 29, 2026 Mistral Medium 3.5 (128B + EAGLE) Mistral AI (France) 128B (dense) Modified MIT License 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 …
April 29, 2026 Ling-2.6-1T inclusionAI / Ant Group (China) 1T MIT 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 … Link
April 29, 2026 Granite 4.1 (3B, 8B, 30B; Base + Instruct) IBM (USA) 3B, 8B, 30B Apache 2.0 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 …
April 28, 2026 Talkie-13B (talkie-1930-13b-base, talkie-1930-13b-it, talkie-web-13b-base) talkie 13B Apache 2.0 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 …
April 28, 2026 SenseNova U1 (8B-MoT + 8B-MoT-SFT) SenseNova / SenseTime (China) 8B MoT Apache 2.0 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 …
April 28, 2026 Nemotron 3 Nano Omni (Reasoning; BF16 + FP8 + NVFP4) NVIDIA (USA) 31B total / 3B active (30B-A3B class MoE) NVIDIA Open Model Agreement 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 …
April 28, 2026 Ling-2.6-flash inclusionAI / Ant Group (China) 104B total / 7.4B active (MoE) Unknown 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. … Link
April 28, 2026 Laguna XS.2 Poolside (USA) 33B total / 3B active (MoE) Apache 2.0 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 … Link
April 27, 2026 MiMo-V2.5 (MiMo-V2.5, MiMo-V2.5-Base, MiMo-V2.5-Pro, MiMo-V2.5-Pro-Base) Xiaomi MiMo (China) MiMo-V2.5: 310B total / 15B active (Sparse MoE); MiMo-V2.5-Pro: 1.02T total / 42B active (MoE) MIT 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 …
April 23, 2026 Hy3-preview (Instruct + Base) Tencent Hy Team (China) 295B total / 21B active (MoE) + 3.8B MTP layer Tencent Hy Community Licens… 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 …
April 23, 2026 DeepSeek-V4 (Flash + Pro; Base + Instruct) DeepSeek AI (China) Flash: 284B total / 13B active (MoE); Pro: 1.6T total / 49B active (MoE) MIT 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 … Link
April 22, 2026 Qwen3.6-27B (Base, FP8) Alibaba Qwen Team (China) 27B dense; FP8 variant released Apache-2.0 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, …
April 20, 2026 Kimi-K2.6 Moonshot AI (China) 1T total / 32B active (MoE) Modified MIT License 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 …
April 15, 2026 Qwen3.6-35B-A3B (Base, FP8) Alibaba Qwen Team (China) 35B total / 3B active (MoE); FP8 variant released Apache-2.0 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 …
April 9, 2026 MiniMax-M2.7 MiniMax (China) ~228.7B total (MoE; 256 experts, 8 active per token; FP8 quantization) Non-commercial MIT-style li… 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, … Link
April 8, 2026 LFM2.5-VL-450M Liquid AI (USA) 450M-class VLM (LFM2.5-350M text backbone + SigLIP2 NaFlex ~86M vision encoder) LFM Open License v1.0 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 … Link
April 7, 2026 GLM-5.1 (Base, FP8) Z.AI / Zhipu AI (China) ~753.9B total (MoE-DSA family; FP8 variant also released) MIT 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 … Link
April 2, 2026 Gemma 4 (E2B, E4B, 26B-A4B, 31B) Google DeepMind (USA) E2B (~5.1B total), E4B (~8B total), 26B-A4B (25.2B total / 3.8B active), 31B dense (30.7B) Apache-2.0 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 … Link
April 1, 2026 Trinity-Large-Thinking Arcee AI ~398B total / ~13B active per token (Sparse MoE, AFMoE) Apache-2.0 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 … Link
March 2026 17 releases
March 31, 2026 LFM2.5-350M Liquid AI 350M (LFM2 hybrid architecture; 16 layers: 10 double-gated LIV conv + 6 GQA) LFM Open License v1.0 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 … Link
March 31, 2026 1-bit Bonsai 8B (GGUF, MLX) PrismML 8.19B (Qwen3-8B dense base; end-to-end 1-bit deployment format) Apache-2.0 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 … Link
March 26, 2026 gpt-oss-puzzle-88B NVIDIA 88B (MoE-derived optimized architecture) NVIDIA Open Model License 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, … Link
March 26, 2026 daVinci-LLM-3B SII-GAIR-NLP ~3.4B (Qwen2-family decoder-only Transformer) Apache-2.0 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 … Link
March 25, 2026 LongCat-Next Meituan LongCat A3B (MoE, native multimodal) MIT 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. … Link
March 23, 2026 Holo3-35B-A3B H Company (France) 35B total / 3B active (Sparse MoE) Apache-2.0 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 …
March 19, 2026 Nemotron-Cascade-2-30B-A3B NVIDIA 30B / 3B active (MoE) NVIDIA Open Model License 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 … Link
March 16, 2026 Nemotron 3 Nano 4B (BF16, FP8, GGUF) NVIDIA 4B (dense/hybrid SLM) NVIDIA Nemotron Open Model … 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 … Link
March 12, 2026 Chroma Context-1 Chroma 20B (MoE, agentic search model) Apache 2.0 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 … Link
March 11, 2026 Nemotron 3 Super 120B-A12B (Base, BF16, FP8, NVFP4) NVIDIA 120B / 12B active (LatentMoE) NVIDIA Nemotron Open Model … 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, … Link
March 9, 2026 MiroThinker-1.7 and MiroThinker-1.7-mini MiroMind AI 235B and 30B (dense, text-generation) Apache 2.0 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 …
March 9, 2026 Covenant-72B 1Covenant 72B (dense) Apache 2.0 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 … Link
March 6, 2026 Sarvam 30B and 105B Sarvam AI (India) 30B / 2.4B active (MoE), 105B / 10.3B active (MoE) Apache 2.0 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, …
March 4, 2026 Yuan3.0 Ultra YuanLabAI 1010B / 68.8B active (MoE) Apache 2.0 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 … Link
March 4, 2026 Phi-4-Reasoning-Vision-15B Microsoft 15B (multimodal reasoning) MIT 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, … Link
March 2, 2026 Step 3.5 Flash Base StepFun 196.81B / ~11B active (MoE) Apache 2.0 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 … Link
March 2, 2026 Qwen 3.5 Small Series (Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B, Qwen3.5-0.8B) Alibaba (Qwen Team) 9B, 4B, 2B, 0.8B (dense) Apache 2.0 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 …
February 2026 15 releases
February 25, 2026 Qwen 3.5 Medium Series (Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, Qwen3.5-27B) Alibaba (Qwen Team) 35B / 3B (MoE), 122B / 10B (MoE), 27B (dense) Apache 2.0 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 …
February 24, 2026 LFM2-24B-A2B Liquid AI 24B / 2.3B active (MoE) LFM Open License v1.0 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 … Link
February 17, 2026 Tiny Aya Cohere Labs (Canada) 3.35B (dense) CC-BY-NC (non-commercial; c… 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 …
February 16, 2026 Qwen3.5-397B-A17B Alibaba (Qwen Team) 397B / 17B (MoE) Apache 2.0 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 …
February 16, 2026 Ling-2.5-1T inclusionAI / Ant Group (China) 1T / 63B (MoE) MIT 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 …
February 12, 2026 Ring-2.5-1T inclusionAI / Ant Group (China) 1T / 63B (MoE) MIT 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 …
February 12, 2026 MiniMax-M2.5 MiniMax (China) 230B / 10B (MoE) MIT (modified, requires 'Mi… 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+ …
February 11, 2026 Nanbeige4.1-3B Nanbeige (China) 3B (dense) Apache 2.0 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 … Link
February 11, 2026 MiniCPM-SALA OpenBMB / ModelBest (China) 9B (dense) Apache 2.0 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 … Link
February 11, 2026 Ming-flash-omni-2.0 inclusionAI / Ant Group (China) 100B / 6B (MoE) MIT 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 … Link
February 11, 2026 GLM-5 Zhipu AI / Z.AI (China) 744B / 40B (MoE) MIT 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 … Link
February 9, 2026 LLaDA2.1-mini inclusionAI / Ant Group (China) 16B (dense) Apache 2.0 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 … Link
February 3, 2026 Qwen3-Coder-Next Qwen Team / Alibaba (China) 80B total / 3B active (MoE) Apache 2.0 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 … Link
February 3, 2026 MiniCPM-o 4.5 OpenBMB / ModelBest (China) 9B Apache 2.0 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 … Link
February 2, 2026 Step 3.5 Flash StepFun (China) 196.81B / ~11B active (MoE) Apache 2.0 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 … Link
January 2026 8 releases
January 28, 2026 OLMo Hybrid 7B Series (Base, Think-SFT, Instruct-SFT, Instruct-DPO, Think) Allen Institute for AI (Ai2) 7B (hybrid RNN/Transformer dense) Apache 2.0 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 … Link
January 27, 2026 Trinity Large (Preview, Base, TrueBase) Arcee AI (USA) 400B total / 13B active (MoE) Apache 2.0 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 … Link
January 27, 2026 SERA-32B Allen Institute for AI (Ai2) - USA 32B (based on Qwen 3-32B) Apache 2.0 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 … Link
January 26, 2026 daVinci-Dev (32B, 72B) GAIR (SII) 32B and 72B (plus MT checkpoints) Qwen license (models); mixe… 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 … Link
January 23, 2026 Mistral Small 4 119B (Base, NVFP4, Eagle) Mistral AI 119B / 6.5B active (MoE) Apache 2.0 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, …
January 7, 2026 AntAngelMed Ant Group / Health Information Center of Zhejiang Province (China) 100B / 6.1B active (MoE, 1/32 activation ratio) Apache 2.0 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 … Link
January 6, 2026 NousCoder-14B Nous Research (USA) 14B (dense) Apache 2.0 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 … Link
January 5, 2026 LFM 2.5 (Base, Instruct, Japanese, Vision-Language) Liquid AI (USA) 1.2B / 1.5B / 1.6B (family variants with hybrid architecture) LFM Open License v1.0 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 … Link
December 2025 24 releases
December 30, 2025 Llama 3.3 8B Instruct (Allura community reconstruction) Allura (Forge) / Fizzarolli (reconstruction of a Meta model) 8,030,261,248 (8.03B; dense, including untied input embeddings and output head) Publisher-declared Llama 3.… 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 …
December 29, 2025 WeDLM-8B WeChat AI, Tencent (China) 8B dense; 8,190,735,360 learned parameters each (Base and Instruct; BF16) Custom Tencent WeDLM terms … 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 … Link
December 29, 2025 HyperCLOVA X SEED 8B Omni NAVER Cloud (South Korea) 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) HyperCLOVA X SEED 8B Omni M… 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 … Link
December 29, 2025 HyperCLOVA X SEED 32B Think NAVER Cloud (South Korea) 32B dense; 33,313,410,304 total checkpoint parameters (32,598,348,800 language model + 688,841,984 vision encoder + 26,219,520 projector) HyperCLOVA X SEED 32B Think… 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, … Link
December 26, 2025 MiniMax-M2.1 MiniMax (China) 230B total / ~10B active (MoE; 228,689,748,992 learned parameters in the released checkpoint) Modified MIT: retain copyri… 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 …
December 25, 2025 LFM2-2.6B-Exp Liquid AI (USA) 2.6B (2,569,272,320 unique learned parameters; dense, tied embeddings) LFM Open License v1.0 (cust… 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 …
December 22, 2025 GLM-4.7 Z.ai / Zhipu AI (China) MoE: nominal 355B total / 32B active; 352.798B main-model learned parameters + 5.540B exported MTP parameters (358.338B packaged, including shared-weight copies) MIT (publisher declarations… 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 …
December 18, 2025 T5Gemma 2 Google (USA) 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 Gemma Terms of Use with the… 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 … Link
December 18, 2025 Hearthfire-24B Latitude (USA) 24B (dense; 23,572,403,200 parameters) Apache 2.0 (publisher decla… 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 …
December 18, 2025 FunctionGemma 270M Google (USA) 270M nominal; 268,098,176 total dense parameters, including 167,772,160 tied embedding parameters Gemma Terms of Use + Gemma … 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 …
December 16, 2025 Nemotron-Cascade NVIDIA (USA) 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) NVIDIA Open Model License A… 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, … Link
December 16, 2025 MiMo-V2-Flash (Base and post-trained) Xiaomi (China) 309B total / 15B active (MoE); 308,778,780,864 main-model parameters + 987,820,224 across three MTP blocks MIT (model weights; GitHub … 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 … Link
December 15, 2025 QwenLong-L1.5-30B-A3B Tongyi Lab, Alibaba Group (China) 30.532B total; approximately 3.353B active per token (MoE, including input/output embeddings) Apache 2.0 (publisher-decla… 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 … Link
December 15, 2025 Nemotron 3 Nano 30B-A3B NVIDIA (USA) 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 Custom NVIDIA Open Model Li… 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 … Link
December 12, 2025 OLMo 3.1 Allen Institute for AI (USA) 32B Think / 32B Instruct / 7B RL-Zero Math and Code (32,233,522,176 / 7,298,011,136 parameters; dense) Apache 2.0 (model weights a… 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 … Link
December 9, 2025 Nomos 1 Nous Research (USA), in collaboration with Hillclimb AI 30.532B total; approximately 3.353B active per token (MoE, including input embeddings and output head) Apache 2.0 (publisher weigh… 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 …
December 9, 2025 Devstral 2 / Devstral Small 2 (2512) Mistral AI (France) 123B nominal Devstral 2 (125,025,988,608 learned parameters); 24B Devstral Small 2 (24,011,361,280 including vision and connector); dense Devstral 2: Modified MIT, r… 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 …
December 8, 2025 GLM-4.6V / GLM-4.6V-Flash Zhipu AI (Z.ai, China) GLM-4.6V: nominal 106B total / 12B active MoE (107.711B stored including vision/projector); Flash: nominal 9B dense (10.293B including vision/projector) MIT (publisher-declared wei… 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, … Link
December 5, 2025 Rnj-1 (Base and Instruct) Essential AI (USA) 8.31B each (8,310,501,376; dense, tied embeddings) Apache 2.0 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 …
December 3, 2025 Hermes 4.3 36B Nous Research (USA) 36,151,104,512 per checkpoint (36B dense; Psyche and centralized variants) Apache 2.0 declared for Psy… 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 …
December 2, 2025 Mistral Large 3 Mistral AI (France) 675B total / 41B active (publisher nominal; granular MoE with a 2.5B vision encoder) Apache 2.0 (explicit weight… 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 …
December 2, 2025 Ministral 3 Mistral AI (France) 3B / 8B / 14B dense family; approximately 3.849B / 8.918B / 13.945B total learned parameters including vision and connectors Apache 2.0 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 … Link
December 1, 2025 Trinity Nano Preview / Mini Arcee AI (USA) Nano Preview: 6.120B total / ~1B active; Mini: 26.124B total / ~3B active (MoE) Apache 2.0 at release; Open… 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 … Link
December 1, 2025 DeepSeek-V3.2 (V3.2 and V3.2-Speciale) DeepSeek AI (China) 671B nominal total / 37B nominal active per token (each MoE); 671.878B main-model tensors plus 13.477B packaged MTP tensors MIT 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 … Link
November 2025 7 releases
November 27, 2025 DeepSeekMath-V2 DeepSeek AI (China) 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 Apache-2.0 for the publishe… 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 … Link
November 26, 2025 INTELLECT-3 Prime Intellect (USA) 106B total / 12B active (MoE; BF16 package contains 106,852,245,504 BF16 parameter elements plus 5,760 FP32 routing-correction elements) MIT (official weight metada… 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 … Link
November 25, 2025 LLaDA2.0 (mini and flash) inclusionAI / Ant Group (China) 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 Apache 2.0 (both checkpoint… 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 … Link
November 21, 2025 Nanbeige4-3B-Thinking-2511 Nanbeige LLM Lab / BOSS Zhipin (China) 3B nominal; 3,933,637,120 total (dense, untied embeddings) Apache 2.0 (model-card decl… 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. … Link
November 20, 2025 OLMo 3 Allen Institute for AI (USA) 7B / 32B dense (7,298,011,136 / 32,233,522,176 parameters) Apache 2.0 (model weights a… 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 … Link
November 12, 2025 Baguettotron Pleias (France) 321M (320,956,992 total; dense, tied embeddings) Apache 2.0 (publisher model… 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 … Link
November 6, 2025 Kimi K2 Thinking Moonshot AI (China) 1T total / 32B active (MoE) Modified MIT; retain notice… 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 …
October 2025 7 releases
October 30, 2025 Kimi Linear 48B A3B Instruct Moonshot AI (China) 48B total / 3B active per token (nominal MoE); 49,122,681,728 checkpoint parameters MIT (explicit Base/Instruct… 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 … Link
October 27, 2025 MiniMax-M2 MiniMax AI (China) 230B total / 10B active (MoE; 228.690B learned parameters in the released main checkpoint) MIT at the October 27 launc… 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 … Link
October 27, 2025 Ming-flash-omni Preview inclusionAI / Ant Group (China) 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 Hugging Face metadata ident… 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 … Link
October 21, 2025 Qwen3-VL-2B / 32B (Instruct / Thinking) Alibaba Qwen Team (China) Dense 2B / 32B; full BF16 checkpoints: 2,127,532,032 / 33,357,390,064 parameters, including vision encoder and four vision-language mergers Apache 2.0 (explicit model-… 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 … Link
October 15, 2025 Qwen3-VL-4B / 8B (Instruct and Thinking) Alibaba Qwen Team (China) 4B / 8B dense language backbones; 4,437,815,808 / 8,767,123,696 total checkpoint parameters including vision encoder and mergers Apache 2.0 (checkpoint decl… 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 … Link
October 13, 2025 Ring-1T inclusionAI / Ant Group (China) 1T total, approximately 50B active per token (MoE) MIT (publisher-declared; no… 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 … Link
October 9, 2025 Ling-1T Ant Group / inclusionAI (China) 1T total MoE (999,705,309,184 learned checkpoint parameters); approximately 50B active per token (technical report: 51B) MIT (publisher hub declarat… 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. … Link
September 2025 10 releases
September 30, 2025 GLM-4.6 Zhipu AI (Z.ai, China) 355B total / 32B active (MoE, nominal); 352.798B main-model learned parameters, 356.786B including packaged MTP MIT (weights); Apache-2.0 (… 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 …
September 29, 2025 Ring-1T-preview Ant Group / inclusionAI (China) Approximately 1T total (MoE); approximately 50B active per token on the Ling-1T backbone MIT (official weight-reposi… 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 …
September 29, 2025 DeepSeek-V3.2-Exp DeepSeek AI (China) 671B backbone / approximately 37B active (MoE; additional DSA indexers; approximately 685B packaged with MTP) MIT 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 … Link
September 23, 2025 Qwen3-VL-235B-A22B Alibaba / Qwen Team (China) 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 Apache 2.0 (checkpoint mode… 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 … Link
September 22, 2025 Qwen3-Omni-30B-A3B (Instruct / Thinking) Alibaba Qwen Team (China) 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 Apache 2.0 (explicit checkp… 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 … Link
September 22, 2025 DeepSeek-V3.1-Terminus DeepSeek AI (China) 671B total / 37B active (MoE; approximately 685B including packaged MTP module) MIT 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 …
September 11, 2025 Qwen3-Next-80B-A3B (Instruct / Thinking) Alibaba (Qwen Team) (China) 80B total / approximately 3B active per token (sparse MoE; 79.674B main-model learned parameters, 81.325B packaged including MTP) Apache 2.0 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 …
September 5, 2025 Kimi K2-Instruct-0905 Moonshot AI (China) 1T total / 32B active per token (MoE; nominal) Modified MIT; retain copyri… 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 …
September 4, 2025 Wayfarer-2-12B LatitudeGames (USA) 12B (12,247,802,880 parameters; dense) Apache 2.0 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 …
September 4, 2025 Nova-70B-Llama-3.3 Latitude (USA) 70B (dense; 70,553,706,496 parameters) Llama 3.3 Community License… 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 …
August 2025 8 releases
August 28, 2025 Command A Translate Cohere / Cohere Labs (Canada) 111B (dense; 111,057,580,032 parameters) CC-BY-NC-4.0 with Acceptabl… 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 … Link
August 26, 2025 Hermes 4 Nous Research (USA) 70B / 405B (dense; exact learned totals: 70,553,706,496 / 405,853,388,800) Llama 3.1 Community License… 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 … Link
August 23, 2025 Grok-2 xAI (USA) 269.515B total / approximately 114.897B active (MoE) Custom Grok 2 Community Lic… 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 …
August 21, 2025 DeepSeek-V3.1 DeepSeek AI (China) 671B total / 37B activated per token (MoE; approximately 685B packaged including auxiliary MTP tensors) MIT 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 …
August 21, 2025 Command A Reasoning Cohere / Cohere Labs (Canada) 111B dense (111,057,580,032 learned parameters) CC BY-NC 4.0 with Cohere La… 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 …
August 20, 2025 Seed-OSS-36B ByteDance Seed Team (China) 36B dense (36,151,104,512 parameters each; Base, Base-woSyn and Instruct) Apache 2.0 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 …
August 14, 2025 Gemma 3 (270M) Google (USA) 270M nominal; 268,098,176 total dense parameters per PT/IT checkpoint (167,772,160 tied embedding parameters + 100,326,016 other parameters) Custom Gemma Terms of Use +… 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 …
August 5, 2025 gpt-oss (gpt-oss-120b / gpt-oss-20b) OpenAI (USA) gpt-oss-120b: 116.83B total / 5.13B active per token; gpt-oss-20b: 20.91B total / 3.61B active per token (MoE) Apache License 2.0; accompa… 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 … Link
July 2025 12 releases
July 31, 2025 Command A Vision Cohere and Cohere Labs (Canada) 112B total (111,867,525,360 checkpoint parameters; 111B text backbone) CC BY-NC 4.0 with Cohere ac… 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, …
July 28, 2025 GLM-4.5 Zhipu AI (Z.ai, China) 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) MIT (weights; explicit laun… 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. … Link
July 22, 2025 Qwen3-Coder-480B-A35B-Instruct Alibaba Qwen Team (China) 480B total (480,154,875,392 learned parameters); 35B active per token; MoE with 160 routed experts, 8 selected, no shared expert Apache 2.0 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 …
July 22, 2025 HyperCLOVA X SEED 14B Think NAVER Cloud (South Korea) 14.748B (14,748,112,896; dense, untied embeddings) HyperCLOVA X SEED 14B Think… 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 … Link
July 21, 2025 Qwen3-235B-A22B-Instruct-2507 Alibaba Qwen Team (China) 235B total / 22B active (MoE; 235,093,634,560 learned parameters) Apache 2.0 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 … Link
July 16, 2025 OpenReasoning-Nemotron NVIDIA (USA) 1.5B / 7B / 14B / 32B (four dense models; checkpoint tensor counts approximately 1.544B / 7.616B / 14.770B / 32.764B) CC-BY-4.0 for all four rele… 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 …
July 15, 2025 Voxtral Mistral AI (France) Mini: 4,674,351,104; Small: 24,259,880,960 learned parameters, including audio encoder and adapter (nominal 3B/24B text-backbone labels; dense) Apache 2.0 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 … Link
July 15, 2025 Ming-Lite-Omni v1.5 inclusionAI / Ant Group (China) 20.3B total, 3B active in the MoE section Hugging Face metadata ident… 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 … Link
July 11, 2025 Kimi K2 Moonshot AI (China) 1T total / 32B active per token (sparse MoE; paper: 1.04T / 32.6B) Modified MIT License (retai… 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. … Link
July 10, 2025 LFM2 Liquid AI (USA) 350M, 700M, 1.2B dense (354,483,968; 742,489,344; 1,170,340,608 unique parameters) LFM Open License v1.0 (cust… 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. … Link
July 9, 2025 T5Gemma Google (USA) 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. Custom Gemma Terms of Use p… 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; … Link
July 8, 2025 SmolLM3-3B Hugging Face (USA) 3B dense (3,075,098,624 unique parameters per checkpoint; tied input/output embeddings) Apache 2.0 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 …
June 2025 3 releases
June 20, 2025 Mistral Small 3.2 Mistral AI (France) 24B dense (24.011B total: 23.572B text + 403M vision + 35.7M multimodal connector) Apache 2.0 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. …
June 16, 2025 MiniMax-M1 MiniMax (China) 456B total; 45.9B active per token (MoE; active count excludes embedding/output matrices) Apache 2.0 (both released M… 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 … Link
June 10, 2025 Magistral Small 1.0 (2506) Mistral AI (France) 24B (23,572,403,200; dense) Apache 2.0 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 … Link
May 2025 5 releases
May 28, 2025 DeepSeek-R1-0528 DeepSeek AI (China) 671B main MoE / 37B activated; approximately 685B packaged including the auxiliary MTP layer MIT (code and model weights… 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; … Link
May 21, 2025 Falcon-H1 Technology Innovation Institute (TII) (UAE) 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 Custom TII Falcon / Falcon-… 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 … Link
May 21, 2025 Devstral Small 2505 (Devstral Small 1.0) Mistral AI (France) and All Hands AI (USA) 24B dense (23,572,403,200 parameters) Apache 2.0 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 … Link
May 16, 2025 Harbinger-24B LatitudeGames (USA) 24B (23,572,403,200 parameters; dense) Apache 2.0 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 …
May 13, 2025 Muse-12B LatitudeGames (USA) 12.248B (12,247,802,880; dense) Apache 2.0 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 …
April 2025 5 releases
April 28, 2025 Qwen3 (original April 2025 release) Alibaba Cloud / Qwen Team (China) 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) Apache License 2.0 (all 14 … 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 … Link
April 24, 2025 HyperCLOVA X SEED NAVER / NAVER Cloud (South Korea) 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) HyperCLOVA X SEED Model Lic… 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 …
April 9, 2025 Kimi-VL Moonshot AI (China) 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 MIT (both original checkpoi… 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 … Link
April 7, 2025 Llama-3.1-Nemotron-Ultra-253B-v1 NVIDIA (USA) 253B (253,401,268,224 parameters; dense) NVIDIA Open Model License +… 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 … Link
April 5, 2025 Llama 4 Meta 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) Llama 4 Community License A… 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 …
March 2025 6 releases
March 24, 2025 DeepSeek-V3-0324 DeepSeek AI (China) 671B total / 37B active per token (MoE; auxiliary MTP weights additional) MIT 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 … Link
March 18, 2025 Llama Nemotron (Nano 8B and Super 49B v1) NVIDIA (USA) Nano: 8B (8,030,261,248); Super: 49B (49,867,145,216); dense, not MoE NVIDIA Open Model License, … 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, … Link
March 17, 2025 Mistral Small 3.1 Mistral AI (France) 24B each (Base and Instruct; dense multimodal; approximately 23.57B text model + 403M vision encoder + 35.65M connector) Apache 2.0 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 …
March 13, 2025 Command A Cohere and Cohere Labs (Canada) 111B dense (111,057,580,032 checkpoint parameters) CC-BY-NC-4.0 with Cohere La… 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 … Link
March 12, 2025 Gemma 3 Google DeepMind 1B / 4B / 12B / 27B (dense language-model sizes; 4B, 12B and 27B additionally use a shared 417M-parameter SigLIP vision encoder) Gemma Terms of Use (custom … 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 … Link
March 4, 2025 Aya Vision Cohere Labs (Canada) 8B and 32B (dense; approximately 8.63B and 33.14B total checkpoint parameters including vision components) CC BY-NC 4.0 with Cohere La… 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 … Link
February 2025 2 releases
February 26, 2025 Phi-4-mini and Phi-4-multimodal Microsoft Research (United States) 3.84B (Phi-4-mini, dense); 5.57B (Phi-4-multimodal, Mixture-of-LoRAs) MIT 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, … Link
February 19, 2025 Wayfarer-Large-70B-Llama-3.3 Latitude (USA), in collaboration with Gryphe Padar 70B dense (70,553,706,496 parameters) Llama 3.3 Community License… 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 …
January 2025 5 releases
January 30, 2025 Mistral Small 3 Mistral AI (France) 23.6B (dense; marketed as 24B) Apache 2.0 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, …
January 20, 2025 DeepSeek-R1 DeepSeek (China) 671B total / 37B active (R1 and R1-Zero MoE); 1.5B, 7B, 8B, 14B, 32B, and 70B (dense distills) MIT (distilled checkpoints … 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 … Link
January 16, 2025 Wayfarer-12B LatitudeGames (USA) 12.248B (12,247,802,880; dense) Apache 2.0 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 …
January 15, 2025 MiniMax-01 MiniMax (China) 456B total / 45.9B active (MoE; Text-01 and VL-01; VL-01 adds a 303M ViT) MiniMax Model License (cust… 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 … Link
January 10, 2025 Sky-T1-32B-Preview NovaSky / UC Berkeley Sky Computing Lab (United States) 32.8B (dense; Qwen2.5-32B-Instruct fine-tune) Apache 2.0 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 …
December 2024 1 release
December 12, 2024 Phi-4 Microsoft Research (United States) 14.7B (dense; marketed as 14B) MIT 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, … Link

Know a model I missed or spot an error? Submit it.