Moonshot AI Opens Kimi K3 Weights, Adds 1M-Token Context and Agent Infrastructure

Moonshot AI has released the full weights for Kimi K3 on Hugging Face and posted its 47-page technical report, model configuration and Kimi K3 License on GitHub, according to Jiemian News. The company also opened MoonEP and AgentENV, while FlashKDA, which had been open-sourced in April, was folded into the full stack.

The model has 2.78 trillion parameters, uses about 104.2 billion per token, and supports a 1,048,576-token context window. Moonshot AI said it created 51,219,741 sandboxes and 1,505,678 image templates during training and evaluation; because sandboxes could spend up to 98% of their lifetime waiting for inference, AgentENV freezes them and restores them in as little as 49 milliseconds.

The license sets commercial terms for qualified MaaS operators: if affiliated entities generate more than $20 million in combined revenue over 12 consecutive months, they must obtain a separate commercial agreement; if a product using K3 exceeds 100 million monthly active users or $20 million in monthly revenue, it must prominently display “Kimi K3.” MoonEP, FlashKDA and AgentENV use the more permissive MIT license.