Moonshot has published Kimi K3.1 Pro’s model weights on Hugging Face, following its earlier release through apps and an API. The November 4 delivery allows third parties to host the model. Its licence permits modification and redistribution but sets additional requirements for large commercial services, making the terms of self-hosting part of the release.
Moonshot AI, the Chinese developer of the Kimi model family, published downloadable Kimi K3.1 Pro weights on Hugging Face on 4 November. The model had already appeared in its applications and programming interface on 23 October, when the company promised the later delivery. The public repository now lets other organizations obtain the weights for their own deployments, separately from access through Moonshot’s hosted services.[1], [2]
The licence permits use, modification, fine-tuning and redistribution. A company selling model access as a service must obtain a separate commercial agreement once its revenue crosses the stated annual threshold. Large products face an additional requirement to display the model’s name when they cross the licence’s user or monthly-revenue thresholds. ArtificialWatch’s comparison of the licence texts found these conditions unchanged from Kimi K3, with the model name updated.[1]
Those additional conditions do not apply to internal enterprise use. The weights were trained in their released numerical format rather than compressed into that format afterward, according to Moonshot’s model card. K3.1 Pro builds on Kimi K3 with changes to attention, the mechanism that selects relevant parts of the input. The release replaces full-attention layers with sparse-attention layers intended to handle long inputs.[1]
