Moonshot releases the open weights of Kimi K3.1 Pro

Moonshot posted on X that it is releasing the model weights of Kimi K3.1 Pro, on Hugging Face as moonshotai/Kimi-K3.1-Pro, and said "we'll keep advancing open intelligence." It had promised the weights by November 4 when it launched the model on October 23.

Community posts describing the release say the model is a mixture-of-experts with 2.8T total and 104B active parameters and a 1M-token context, that the weights are about 1.56 TB in MXFP4, that vLLM and SGLang can serve it, and that it ships under a custom "Kimi K3.1 Pro License". These details come from users' posts, not from Moonshot's post, which carries only the link.

The SGLang project says it had day-0 support and has added Kimi K3.1 Pro recipes to its Cookbook. It says that on GB300 NVL72 it delivers 420+ tokens/s per user at low concurrency and 31K+ tokens/s per GPU at high concurrency, with B300, GB200 NVL72, B200 and MI355X results also available. These figures are SGLang's own.

Arena, which runs the Agent Arena leaderboard, says that across more than 17,500 real-world, long-horizon agentic sessions Kimi K3.1 Pro (Max) ranks #1 among open models and #13 overall, and #1 among open models in its Work, Code and Chat categories. Arena puts the net improvement at +5.1%, 4.2 percentage points above Kimi K3 (Max). Among open models it ranks #1 in Bash Reliability, Praise vs Complaint and Steerability, and #4 in Confirmed Success.