❯ Moonshot releases Kimi K3.1 Pro weights, bringing its million-token model to self-hosting

From hosted access to downloadable weightsMoonshot released Kimi K3.1 Pro weights on November 4 through its official Hugging Face repository. The model was already available through Kimi’s web and mobile products, Kimi Work, Kimi Code and its API. The release extends access to infrastructure controlled by the user.

Built for sustained tasksThe official technical introduction describes a natively multimodal, three-trillion-class agentic model with a 1-million-token context window, allowing large volumes of text and related material to remain in context. Expanded post-training on Kimi K3 targets coding, 3D creation and financial research. Examples include writing and executing scripts in Blender or CAD tools to build assets from sketches, reference images and text instructions.

Deployment support followsThe repository lists deployment recipes for vLLM, SGLang and TokenSpeed, alongside an API compatible with OpenAI and Anthropic calling conventions. Downloadable weights let organizations choose their data flows, hardware and serving capacity, while taking responsibility for operating the model.

Commercial licensing thresholdsThe Kimi K3.1 Pro License permits deployment and modification. A company operating a model-as-a-service business must obtain a separate agreement before commercial use if combined revenue with affiliates exceeds $20 million over any consecutive 12 months. The terms exempt purely internal use, creating different arrangements for model API providers and internal enterprise deployments.

▮ SIGNALOpen weights turn model capability into an infrastructure choice, making serving efficiency and deployment support part of product competition.