Moonshot AI releases Kimi K3.1 Pro weights
Original: Kimi K3.1 Pro model weights are out
Why This Matters
Open pro-tier multimodal weights raise the bar for self-hosted AI deployment options.
Moonshot AI has published the model weights for Kimi K3.1 Pro on Hugging Face. The multimodal model supports image-text-to-text tasks and is compatible with Transformers, vLLM, SGLang, and Docker deployment.
Moonshot AI has open-released the weights for Kimi K3.1 Pro on Hugging Face, where the model card has already drawn 482 likes. The release covers 8-bit precision weights in Safetensors format under the kimi-k3.1-pro license.
Kimi K3.1 Pro is a multimodal model handling image-text-to-text generation. Deployment options are broad: Transformers, vLLM, SGLang, Docker Model Runner, and an OpenAI-compatible API endpoint out of the box. Quantized variants are available for llama.cpp, Ollama, and LM Studio.
The model card references benchmark results across coding, agentic tasks, reasoning, knowledge, multimodal, and cybersecurity evaluations — though specific scores were not included in the available excerpt. Community access via HuggingChat is also listed as an inference option.
Moonshot AI, based in China, has been building the Kimi series as a direct competitor to frontier models. Releasing pro-tier weights publicly is a notable step, putting capable multimodal reasoning in the hands of developers who can self-host.