Moonshot opens Kimi K3.1 Pro weights for self-hosting
The 2.8-trillion-parameter multimodal model has 104 billion active parameters, a million-token context window and commercial-use conditions.
Moonshot AI has released the model weights for Kimi K3.1 Pro, turning its latest frontier system from a hosted product into a model that teams can deploy on their own infrastructure. The public Hugging Face repository describes it as a natively multimodal, agentic mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters and a context window of just over one million tokens.
The model accepts text and images and is designed for coding, financial research, 3D creation and other long-running agent tasks. Moonshot recommends vLLM, SGLang or TokenSpeed for deployment. The weights use MXFP4 quantization, while activations use MXFP8, which helps reduce the cost of serving a model of this scale.
Moonshot reports competitive results against leading closed models on several coding and agent benchmarks. Those numbers are vendor-reported, however, and some tests are internal or use different agent harnesses, so they should not be read as a neutral ranking.
The release is open-weight rather than fully open source. The custom license permits use, modification and redistribution, but model-as-a-service businesses earning more than 20 million US dollars over 12 months need a separate commercial agreement. Products above 100 million monthly users or 20 million US dollars in monthly revenue must also display the Kimi K3.1 Pro name. Internal use is exempt.
For developers, the practical change is control: the same large model can now be inspected, fine-tuned and hosted outside Moonshot’s API, provided teams can supply the substantial compute it requires.
