560B total / 25B active (572.1B serialized; BF16 base and blockwise-FP8 post-trained checkpoints)
Up to 1M tokens advertised (262,144-token checkpoint configuration)
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 as Ling-3.1-flash: 72 transformer layers arranged as 63 Kimi Delta Attention layers and nine Gated Multi-head Latent Attention layers in a repeating 7:1 pattern, four initial dense layers, 512 routed experts selecting eight per token, one shared expert, 64 attention heads, a 4,096-dimensional hidden state, 1,280-dimensional routed and shared experts, 10,240-dimensional dense feed-forward layers, a 157,184-token vocabulary, and one NEXTN multi-token-prediction layer. inclusionAI markets the architecture as 560B total and 25B active parameters, while Hugging Face inventories 572,120,680,768 serialized parameters. The base was trained using FP8 but released as BF16 after Warmup-Stable and Merge checkpoint merging; its 109 safetensor shards occupy approximately 1.144 TB. The post-trained 1019 checkpoint uses blockwise FP8 E4M3 weights with dynamic activation quantization and contains 54 mixed FP8/BF16 safetensor shards totaling approximately 573.8 GB. The base and post-trained configurations share the same principal architecture and 262,144-position limit, but the base omits the post-trained checkpoint's FP8 quantization declaration and uses a different end-of-sequence token identifier. inclusionAI continues to advertise operation up to one million tokens despite the shipped 262,144-position configurations. The base model is evaluated on 47 benchmarks spanning language, knowledge, code, mathematics, reasoning, multilingual capability, medicine, and long-context understanding. Publisher-reported base scores include 76.71 on MMLU-Pro, 56.57 on GPQA Diamond, 79.88 on HumanEval-Plus, 70.55 on MultiPL-E, 79.40 on MATH500, 60.24 on OlympiadBench, 90.25 on BBH, 87.66 on MultiGSM, 51.90 on LongBench, and 86.64 on RULER at 256K. For the post-trained 1019 checkpoint, inclusionAI reports that additional post-training raises the Artificial Analysis Intelligence Index v4.3.2 from 41 to 44, AA-Briefcase from approximately 1,400 to 1,524 Elo, AutomationBench-AA from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. Thinking remains enabled by default but can be disabled, while structured tool calling, reasoning parsing, expert parallelism, prefix caching, and NEXTN speculative decoding are supported through SGLang and vLLM. The publisher does not disclose the base model's pretraining corpus or token budget, nor the 1019 checkpoint's additional post-training data, optimization method, token budget, or compute. The 1019 model card's SGLang example still names the predecessor repository, so deployers must substitute the 1019 path.