Moonshot publishes the full Kimi K3.1 Pro weights
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
One of the strongest coding and agent models can now be fine-tuned and deployed on private infrastructure, although its size makes that practical mainly for large teams.
Before the weights were published, Kimi K3.1 Pro was available through its API and apps. The MoE model has 2.8 trillion total parameters and 104 billion active parameters per token, a one-million-token context and text and image inputs. The repository is open without an access request. Weights are stored in MXFP4, activations are computed in MXFP8 and quantisation was accounted for during training. According to an independent Vals AI measurement cited by Habr, the model scored 63.64% on Terminal-Bench 4.0, compared with Moonshot's claimed 65.7%, and outperformed DeepSeek V4.1 Flash on the aggregate Vals Index. It nevertheless trailed its predecessor on certain legal and tax tasks. The weights use the custom Kimi K3.1 Pro licence, while local deployment requires cluster-scale infrastructure.
Sources 2