Moonshot AI · Open agentic MoE

Kimi K3.1 Pro

Kimi K3.1 Pro is an open-weight, natively multimodal agentic model from Moonshot AI, described as its most capable 3T-class model to date. It is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and is the first model in the K3.1 series. The card claims improvements in coding, 3D creation, financial research and factual accuracy, with a 1-million-token context window. It is a Mixture-of-Experts model with 2.8T total and 104B activated parameters, 896 experts (16 selected per token), and a 401M-parameter MoonViT-V2 vision encoder. Compared with K3, the full-attention Gated MLA layers are replaced by Gated DeepSeek Sparse Attention for more efficient long-context modelling, and it uses MXFP4 weights with MXFP8 activations via quantization-aware training. Reported results use maximum reasoning effort: Terminal-Bench 4.0 65.7, SWE Marathon v1.1 51.7, BrowseComp 92.1, GPQA-Diamond 94.2 and HLE-Full 51.1. Scores are mixed against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. The model was created on the Hugging Face hub on 2 November 2026 under a custom "other" licence (kimi-k3.1-pro) for use with transformers.