The company says it is available immediately on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Kimi says K3 uses Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. Its stated inference configuration activates 16 of 896 experts. The launch defaults to a maximum thinking-effort setting; low- and high-effort modes are planned for later updates. Kimi says full model weights will be released by July 27, 2026.
The announcement reports company-run evaluations in long-horizon coding, GPU-kernel optimization, compiler development, research workflows, and chip design. One reported 48-hour autonomous chip-design run used open-source EDA tools and a Nangate 45nm library. Kimi also reports a roughly 2.5x scaling-efficiency improvement over Kimi K2.
The source does not provide independent validation of its benchmarks or case studies. It does not establish the final licensing terms for the weights, nor does it show that the full weights are publicly available before the stated release date.