Moonshot releases Kimi K3.1 Pro as open weights: 2.8T parameters, a million-token context, and agentic scores that split against closed flagships
Synopsis
Moonshot AI has published Kimi K3.1 Pro on Hugging Face as an open-weight, natively multimodal agentic model with 2.8 trillion total parameters, 104B activated per token, and a context length of about 1.05 million tokens. It is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and the model card reports substantial gains in coding, 3D creation, financial research, factual accuracy and other agentic capabilities. The card compares it benchmark by benchmark at maximum reasoning effort against Claude, GPT-6 Astra and DeepSeek models, leading on some tasks and trailing on others.
Interpretation
The open-weight Kimi K3.1 Pro pushes total parameters to 2.8 trillion with 104B activated, extends the context window to about 1.05 million tokens, and natively accepts text and images. The model card calls it the first model in the K3.1 series and Moonshot's most capable 3T-class model to date, post-trained from Kimi K3. The card lists a mixture-of-experts architecture with 93 layers, 896 experts, 16 selected experts plus 2 shared experts per token, and a 401M-parameter MoonViT-V2 vision encoder.
On agentic benchmarks, Kimi K3.1 Pro scores 70.0 on HLE-Full with tools, 92.1 on BrowseComp and 43.2 on Agents' Last Exam (ALL). In the same table, Claude Fable 5.1 scores 65.6 on HLE-Full with tools, GPT-6 Astra scores 91.5 on BrowseComp and 41.8 on Agents' Last Exam (ALL). These scores use reasoning effort max, temperature 1.0 and top-p 0.95, while comparison models are marked max with fallback.
Coding results are mixed: Terminal-Bench 4.0 reaches 65.7, above Claude Fable 5.1 at 57.9, but Terminal-Bench-Science 0.1 reaches 38.1, below GPT-6 Astra at 68.1 and Claude Opus 5.5 at 63.3. In the same table, Program Bench gives Kimi K3.1 Pro 31.5 against Claude Opus 5.5 at 65.0 and Claude Fable 5.1 at 53.5. Terminal-Bench 4.0 is run with the Claude Code harness, and the remaining scores come from model blog posts or official leaderboards.
Long-context modeling swaps the full-attention Gated MLA layers of Kimi K3 for Gated DeepSeek Sparse Attention (Gated DSA) layers. The model card describes the replacement as a route to more efficient long-context modeling. The attention stack comprises 69 KDA layers and 24 Gated DSA layers, with a DSA index head dimension of 128, 32 index heads and top-K of 2048.
Perspective
The model card reports a set of benchmark scores at reasoning effort max, temperature 1.0 and top-p 0.95, spanning coding, agentic, reasoning and knowledge, multimodal and security tasks, compared against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. For deployment it gives recipes for the vLLM, SGLang and TokenSpeed inference engines and an OpenAI- and Anthropic-compatible API.
The model card notes that some comparison scores come from other vendors' blogs or official leaderboards, so harnesses and budgets are not uniform; PostTrainBench, for instance, is evaluated internally on H20 GPUs with DeepSeek-V4.1-Flash-1017 as the judge model. A dash marks a score the source table does not report, and Elo-style scores are not directly comparable with accuracy or pass rates. The card also states that proprietary-model testing may involve fallback behavior.
