Moonshot AI has released Kimi-K3, a large multimodal model on Hugging Face, marking the Chinese AI lab's latest push into open-source model distribution.

The model features 2.8 trillion parameters and supports image-text-to-text tasks through a conversational interface. Kimi-K3 uses 8-bit precision quantization and includes compressed tensor optimization for efficient deployment.

The release includes integration with multiple inference providers including Together AI, Fireworks AI, Baseten, and DeepInfra. The model supports both vLLM and SGLang serving frameworks for production deployment.

Technical specifications

Kimi-K3 processes both text and image inputs simultaneously, enabling users to ask questions about visual content. The model card shows evaluation results with a [email protected] score and mean reward of 37.8 on the LHTB Harbor benchmark.

The model requires custom code execution and uses Safetensors format for secure weight storage. Total file size reaches 1.56TB across sharded model files, reflecting the substantial parameter count.

Hugging Face lists the model under a custom "kimi-k3" license rather than standard open-source terms. The model supports feature extraction tasks beyond conversational applications.

Moonshot AI previously gained attention for its Kimi chatbot's long-context capabilities. The company competes with other Chinese AI labs including Zhipu AI and 01.AI in the foundation model space.

The release comes as Chinese AI companies increasingly distribute models through Hugging Face to reach global developer audiences. Kimi-K3 joins other multimodal models available on the platform from both Chinese and Western labs.

Developers can access the model through Hugging Face's Transformers library or deploy it using Docker containers with GPU support. The model card includes example prompts for image analysis and general conversation tasks.