Kimi K3.1 Pro goes open-weight: 2.8T params, free to self-host

Moonshot AI has published the Kimi K3.1 Pro weights on Hugging Face, and it is a huge download. The mixture-of-experts model has 2.8T total parameters, 104B active per token, and a 1M-token context window. It reads images through a built-in vision encoder. The weights come in MXFP4, trained with quantization in mind, so running it yourself takes a multi-GPU server rather than a gaming rig.

It scores close to the top closed models. On Terminal-Bench 4.0 it gets 65.7, against 66.4 for Claude Opus 5.5. Its own model card also lists 92.1 on BrowseComp and 94.2 on GPQA-Diamond. The API, kimi-k3.1-pro, costs $2/$10 per million input/output tokens.

The interesting part is the cyber trade-off. Moonshot says it removed cyber-related data from training for the open release, because it does not want to "extend the absolute frontier of cyber capability" with open weights. On its internal exploit benchmark the score fell from 90.5% to 81%. The Kimi K3.1 Pro License is MIT-style, but any company selling the model as an API service with over $20M in revenue needs a separate deal with Moonshot.

Why it matters: frontier-class coding agents can now run on your own hardware, and "nerf the hacking skills first" may become the standard way to release open weights.