Moonshot AI has officially uploaded the complete model weights of Kimi K3 to Hugging Face, with the technical report simultaneously published on GitHub. This model, featuring 2.8 trillion parameters, 104 billion activations, and support for a 1 million token context, has become the world's first open-weight model to enter the 3 trillion parameter level - exactly 11 days after it was first released in the form of a managed API on July 16.

Within these 11 days, the story has long gone beyond the technical realm: K3 scored third globally (57 points) in the Artificial Analysis intelligence index, behind only Claude Fable5 and GPT-5.6 Sol; it topped the front-end programming arena Arena with 1679 points; Hong Kong's Z.ai stock plummeted 30% during the day, and MiniMax dropped 16%; Moonshot AI's daily revenue surged by at least six times, and its valuation anchor skyrocketed from 20 billion dollars to 50 billion dollars; the White House Office of Science and Technology Policy Director publicly accused it of distilling Anthropic technology, while the Secretary of the Treasury warned of sanctions.

The license is the real "hidden weapon"

Along with the weights, the Kimi K3 License appears to be under the MIT license, but actually contains hidden commercial rules. The agreement stipulates that any entity providing K3 inference or fine-tuning services to third parties, with annual revenue exceeding 20 million dollars, must sign another agreement with Moonshot AI; commercial products with over 100 million monthly active users or monthly revenue exceeding 20 million dollars must prominently label "Kimi K3" on their interface. This is the first time a leading Chinese model company has explicitly included the practice of "cloud vendors hitchhiking" as a priced commercial clause in an open-source license - those using it purely internally or accessing through official certified inference partners are not subject to these restrictions.

In terms of pricing, the official API price for K3 is 3 dollars per million input tokens and 15 dollars per million output tokens, while Claude Fable5 is 10 dollars per million input tokens and 50 dollars per million output tokens. However, multiple testers have reported that K3 consumes more tokens for the same task, and with the default inference mode enabled, the actual cost advantage per task is not as significant as the pricing table suggests.