Source: Hugging Face: moonshotai/Kimi-K3.1-Pro model card
Moonshot releases open weights for Kimi K3.1 Pro
Why it matters
Moonshot has published the full weights of Kimi K3.1 Pro on Hugging Face, keeping the Nov 4 date it promised at launch on October 23. You can now download and self-host the same multimodal agentic model that has been on the Kimi API, under a permissive custom licence with conditions for very large companies.
The model is a 2.8T-parameter mixture-of-experts model with 104B active parameters, a 1M-token context and text plus image input, shipped as MXFP4 weights; Moonshot recommends serving it with vLLM, SGLang or TokenSpeed, so plan for multi-GPU server hardware, not a workstation. Set thinking with the reasoning_effort field (low, high or max, which is the default), and in multi-turn or tool-calling code pass back the full assistant message, including reasoning_content and tool_calls, or the model loses its own reasoning.
The Kimi K3.1 Pro License lets you use, modify, fine-tune and sell the model, but a Model as a Service business with over $20M revenue in 12 months needs a separate agreement with Moonshot, and products with over 100M monthly users or $20M monthly revenue must show "Kimi K3.1-Pro" in the interface; internal use and use through Moonshot or its certified partners are exempt.
That is not an OSI open-source licence. The benchmark table on the model card (for example 65.7 on Terminal-Bench 4.0, run in the Claude Code harness) is Moonshot's own; Artificial Analysis's independent listing is linked below. If you were waiting on the weights to test Kimi K3.1 Pro in your own stack, you can start now; if you only need the API, nothing changed.