Chinese AI lab Moonshot has published the model weights and technical report for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model, under a permissive open-weight license.

K3 activates about 104B parameters per token and adds native vision, a 1M-token context window and roughly 2.5x better compute efficiency than its predecessor Kimi K2, driven by a new attention design.

Moonshot also released supporting infrastructure, including its attention kernels and inference recipes. K3 is billed as the largest open-weight model to date and rivals top frontier US systems, though early coverage notes it still trails the strongest closed models. The weights are open, but training data and code are not.