Moonshot AI has published the Kimi K3.1 Pro model weights on Hugging Face. The release completes the weight-publication step previously planned for November 4 after the model became available through products and an API.
What open weights provide
Downloadable weights let teams with suitable infrastructure run and adapt the model in their own environments, subject to its license. The model card describes text and image input, a context window of roughly one million tokens and a mixture-of-experts architecture.
Moonshot lists 2.8 trillion total parameters and 104 billion activated parameters per token. These figures describe different things: the full set of model parameters versus the portion used for an individual token.
Hardware and license
Model size matters when planning a local deployment. AMD documents one server configuration using eight Instinct MI355X GPUs and the released MXFP4 weights. It is an example, not proof that this is the only possible setup or that an ordinary desktop can run the full model.
Moonshot’s model card links the Kimi K3.1 Pro license. Review its terms and your own memory requirements before downloading or incorporating the model into commercial work.
