Moonshot Publishes Kimi K3.1 Pro Weights on Hugging Face
Moonshot AI has published the Kimi K3.1 Pro weights on Hugging Face, meeting the November 4 date it set when the model launched. The checkpoint holds 2.8 trillion parameters with 104 billion activated per token and a 1M-token context window kept manageable by Gated DeepSeek Sparse Attention. Moonshot’s published benchmarks put it ahead of GPT-6 Astra on all three agentic tests and ahead of Claude Opus 5.5 on two: SWE Marathon 51.7 against 46.3 and 42.5, BrowseComp 92.1 against 91.5 and 90.8, and Terminal-Bench 4.0 65.7, which trails Opus 5.5 at 66.4.
The model handles images, coding and research tasks. Weights ship in MXFP4 and run under vLLM or SGLang. The files total roughly 1.56 TB, so local deployment needs real hardware.
Related: Moonshot Releases Kimi K3.1 Pro With 1M Context, Weights Due by November 4, Moonshot AI Releases Kimi K3 with 2.8 Trillion Open Weights