Moonshot AI released the open weights and technical report for Kimi K3, a 2.8 trillion parameter mixture-of-experts model featuring a 1 million token context window and native vision capabilities. The model activates 16 of 896 experts per token and utilizes a Kimi Delta Attention architecture that delivers 2.5 times greater intelligence per unit of compute compared to its predecessor.

Alongside the model, Moonshot open-sourced FlashKDA attention kernels and the MoonEP communication library to accelerate large-scale training and inference. The release immediately enabled day-zero deployment support from inference providers including vLLM, SGLang, Modal, and Nebius, while the model topped Hugging Face trending within 30 minutes. The weights are distributed under a modified MIT license that requires larger hosting companies earning over $20 million in annual revenue to secure a separate agreement.