Beijing-based Moonshot released the full weights, technical paper, and AgentENV training system for its Kimi K3 artificial intelligence model. The model operates on a 2.8T parameter mixture of experts architecture, supports a 1M token context window and includes native vision capabilities.

Inference engines and cloud providers quickly integrated the model for deployment. vLLM offered day zero support for the architecture, while Red Hat released a quantized checkpoint optimized for Nvidia Hopper processors. Perplexity also added Kimi K3 to its Pro and Max subscription tiers, hosting the model on servers located in the United States.