Moonshot AI opens the weights of Kimi K3.1 Pro: a 2.8-trillion-parameter MoE with a 1M-token context
According to its Hugging Face model card, Kimi K3.1 Pro is an open-weight, natively multimodal (text and image) agentic model, post-trained from Kimi K3 with large-scale reinforcement learning across hundreds of domains. It is a Mixture-of-Experts model with 2.8 trillion total parameters, 104 billion active per token, 93 layers, 896 experts (16 selected per token) and a context window of 1,048,576 tokens; compared with K3, the full-attention Gated MLA layers are replaced by Gated DeepSeek Sparse Attention layers for more efficient long-context modelling. Weights use MXFP4 (quantization-aware training) and the vision encoder, MoonViT-V2, has 401 million parameters. Moonshot reports 65.7 on Terminal-Bench 4.0 (Claude Opus 5.5: 66.4), 92.1 on BrowseComp, 94.2 on GPQA-Diamond and 70.0 on HLE-Full with tools; most Kimi results were measured with the Claude Code or Kimi Code harness. The model uses its own Kimi K3.1 Pro License and is available through the API at platform.kimi.ai. According to Arena's post and IT Home (Nov 5, 2026), on the Agent Arena leaderboard (more than 17,500 real long-horizon agent sessions) Kimi K3.1 Pro (Max) reaches a net improvement of +5.1% (K3: +0.9%), first among open-weight models and 13th among all models evaluated.
Source: Hugging Face / Moonshot AI
Need an LLM chatbot or agent for your business?
Describe your need briefly; you will get clarifying questions and a scoped proposal before any work starts.
Related news
Kakao unveils Canana Space, Canvas and AI Crew: bringing AI agents into KakaoTalk's 50 million users
Tencent releases Hy4: first vision and computer-use support, aimed at the top tier of open models
