Moonshot AI Releases Kimi K3.1 Pro Open-Weight Multimodal Agentic Model

Moonshot AI has released Kimi K3.1 Pro, an open-weight, natively multimodal agentic model and its most capable 3T-class model to date. Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, it is the first model in the K3.1 series and brings a 1-million-token context window. The company says it delivers substantial improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities.

Architecture and specifications

Kimi K3.1 Pro uses a Mixture-of-Experts architecture with 2.8 trillion total parameters and 104 billion activated parameters. It has 93 layers, one dense layer, and an attention-layer composition of 69 KDA layers plus 24 Gated DSA layers. The context length is 1,048,576 tokens, the vocabulary size is 160K, and the attention hidden dimension is 7,168 with 96 attention heads. Compared with Kimi K3, full-attention Gated MLA layers are replaced with Gated DeepSeek Sparse Attention layers for more efficient long-context modeling. The model uses SiTU-GLU activation, a MoonViT-V2 vision encoder with 401 million parameters, and MXFP4 weights with MXFP8 activations through quantization-aware training. It has 896 experts, selects 16 per token, and includes two shared experts; the latent MoE dimension is 3,584 and each expert has a hidden dimension of 3,072. DSA index settings include a head dimension of 128, 32 heads, and top-k of 2,048.

Evaluation results

Unless otherwise specified, reported results use reasoning effort set to max, temperature 1.0, and top-p 0.95. A dash in the source table means no reported score. Higher scores are better except where a benchmark explicitly states lower is better, and Elo scores are not directly comparable with accuracy or pass-rate scores.

Kimi K3.1 Pro records 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, 70.0 on HLE-Full with tools, 94.2 on GPQA-Diamond, 79.6 on OSWorld V2.1, 1846 Elo on GDPval-AA v2.1, and 8.3 on Kimi Hallu Bench Hard with tools, where lower is better. It also records 75.5 on DeepSWE v1.1, 92.1 on BrowseComp, 90.0 on MMMU-Pro with python, 69.4 on APEX-Agents, 77.1 on the internal Kimi Code Bench 2.0, and 61.2 on the internal Kimi Code Bench 3.0 Long-horizon benchmark.

Deployment and availability

The API is available at platform.kimi.ai by selecting kimi-k3.1-pro, with OpenAI- and Anthropic-compatible APIs. Recommended inference engines are vLLM, SGLang, and TokenSpeed. The model is released under the Kimi K3.1 Pro License, and questions can be sent to [email protected]. The Hugging Face page lists 797 downloads last month, Safetensors with model size 2.8 trillion parameters, tensor types F32, BF16, and U8, two finetuned models in the model tree, and one Space using the model.

Model usage

Thinking effort is configured with the top-level reasoning_effort request field, which supports low, high, and max, with max as the default. For multi-turn conversations and tool calls, the complete assistant message returned by the API must be passed back to messages as-is, including reasoning_content and tool_calls, not just content. Full guides and examples cover vision input, structured output, partial mode, tool choice, dynamic tool loading, and context caching.

Sources: huggingface.co