Open source · 27 July 2026

Moonshot opens up Kimi K3, a 2.8T-parameter multimodal MoE model

Moonshot AI published weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with about 104B active parameters, a 1M-token context window and native text, image and video input. The weights ship with MXFP4 quantization-aware training to ease deployment, under a custom Kimi K3 license. For practitioners, it was widely reported as the largest open-weight release to date, though serving it still requires multi-node GPU clusters.