The launch
China's Moonshot AI put Kimi K3 in its API and apps on July 16, 2026 and promised weights by month end. On July 27, at "Kimi K3 Open Day", it published the weights, a technical report and three infrastructure projects on GitHub (repository MoonshotAI/Kimi-K3).
Key numbers
- 2.8 trillion parameters in a Mixture-of-Experts design, with 16 of 896 experts active per token.
- 1-million-token context window and native text, image and video input.
- New techniques: Kimi Delta Attention (KDA) and Attention Residuals, which the company says deliver about 2.5x scaling efficiency over Kimi K2.
- MXFP4 quantization-aware training from SFT onward, for broader hardware support.
- Open-sourced alongside: MoonEP (MoE communication), FlashKDA (kernels) and AgentEnv (agent sandboxes).
In Moonshot's own benchmarks, K3 sits close to top closed models and leads some agent tests such as BrowseComp and MCPMark. Those figures come from the company and still need independent verification.
Caveats
The weights ship under the Kimi K3 License, not a standard license like MIT or Apache. Commercial use outside the covered cases requires reading the terms. And serving a model this size needs a high-end GPU cluster, so "open" does not mean "cheap to run".
Why it matters
According to Moonshot, it is the first open model in the 3-trillion-parameter class, and it reinforces Chinese labs' push for competitive open weights.