Kimi K3.1 Pro, released with open weights by Moonshot AI on 11 2: It is the first model in the K3.1 series, trained via large-scale joint reinforcement learning post-training on Kimi K3. It features an MoE architecture with 2.8T total parameters and 104B active parameters per token, natively supports text and image inputs (MoonViT-V2 vision encoder), and has a context length of 1,048,576 tokens; weights are quantized to MXFP4 via quantization-aware training (activations are MXFP8). Compared to K3, it replaces full-attention Gated MLA layers with Gated DSA (DeepSeek Sparse Attention), resulting in greater efficiency for long contexts. Officially, significant improvements are claimed in coding, 3D creation, financial research, factual accuracy, and agentic capabilities. Not included in the recent 7 periods.
Rising score 499 · Likes 510 · Downloads 797
Reasons
- In official benchmarks (max reasoning intensity), leading scores are mostly in long-horizon coding and agentic tasks: SWE Marathon v1.1 51.7, HLE-Full with tools 70.0, Agents' Last Exam 43.2, and OfficeQA Pro v2 64.4 are all the highest in the table; Terminal-Bench 4.0 65.7 and OSWorld V2.1 79.6 are close to Claude Opus 5.5's scores of 66.4 / 81.8; BrowseComp 92.1.
- Officially claimed to have fewer hallucinations: On the internal Kimi Hallu Bench Hard (with tools, lower is better), it scored 8.3, with only GPT-6 Astra (5.5) scoring lower in the table, while DeepSeek-V4.1-Flash-1017 scored 23.3.
- Architecture and evaluation settings are publicly detailed: 93 layers (69 KDA layers + 24 Gated DSA layers), 896 experts with 16 selected per token plus 2 shared experts, vocabulary size of 160K; the agent framework used for each benchmark (mostly Claude Code or Kimi Code) and the sources for comparison scores are noted in footnotes.
Usage
- Without self-hosting: Select kimi-k3.1-pro on platform.kimi.ai; official OpenAI / Anthropic compatible interfaces are provided.
- Self-hosting: Official recommendations include vLLM, SGLang, and TokenSpeed, with deployment recipe links for each provided in the card (e.g., recipes.vllm.ai/moonshotai/Kimi-K3.1-Pro); the repository includes custom code (configuration_kimi_k3.py, kimi_k3_processor.py, etc.).
- Reasoning intensity is set via reasoning_effort at the top level of the request to low / high / max (default is max); for multi-turn conversations and tool calls, the entire assistant message returned by the API must be placed back into messages exactly as is, including reasoning_content and tool_calls; official evaluations use temperature 1.0 and top_p 0.95.
Risks
- All scores are self-reported by the official team, with 9 items being internal benchmarks (Kimi Code Bench, Business Operation, PerceptionBench, Kimi Hallu Bench, etc.); scores for closed-source models may include regressions; there are also clearly lagging items: Program Bench 31.5 (Claude Opus 5.5 scored 65.0), Terminal-Bench-Science 0.1 38.1 (GPT-6 Astra 68.1), and HLE-Diamond 39.4 (GPT-6 Astra 65.8).
- Extremely large scale: Hardware requirements are not specified in the card; with 2.8T total parameters, even using MXFP4, it can only run on multi-GPU server clusters, meaning individual users are essentially limited to API access; it only supports text and image inputs.
- The license is a modified MIT (Kimi K3.1 Pro License): If operating as a Model-as-a-Service provider and combined revenue with affiliates exceeds 2,000 million USD over 12 months, a separate agreement with Moonshot AI must be signed before commercial use; if monthly active users exceed 1 billion or monthly revenue exceeds 2,000 million USD, Kimi K3.1-Pro must be prominently displayed in the interface; internal use is exempt from these two restrictions.