Kimi K3 is a 2.8 trillion parameter open-weight AI model from the Chinese lab Moonshot AI, released on 26 July 2026. It is the largest open-weight model published so far, and anyone can download the weights and run it. Underneath the huge headline number it is a sparse mixture of experts, so it activates only 16 of its 896 expert sub networks for any single request. It ships with a one million token context window and native support for images, and early third party tests rank it first for building web interfaces and second overall behind Anthropic's Claude Fable 5.
What is Kimi K3?
Kimi K3 is Moonshot AI's newest flagship, and its most open. "Open weight" means the trained model itself is published for download, not just rented through a paid service, so you can host it on your own hardware, inspect how it behaves, and build on it without asking permission. Moonshot released the weights on 26 July 2026, roughly a day ahead of the 27 July date it had previously signalled. That timing matters, because an open model of this size changes what individuals and small teams can run without a cloud contract.
Why does 2.8 trillion parameters not mean 2.8 trillion in use?
The headline number is the total size of the model, but Kimi K3 does not fire all of it at once. It is a sparse mixture of experts, so for each request it routes the work to just 16 of its 896 expert sub networks. That keeps the running cost far closer to a much smaller model while still drawing on the knowledge packed into the full 2.8 trillion parameters. A one million token context window lets it hold very long documents or codebases in view at the same time, and Moonshot points to a technique it calls Kimi Delta Attention for keeping long conversations efficient.
