Moonshot publishes the Kimi K3 weights under its own licence

Moonshot AI/Kimi models

Moonshot published Kimi K3's weights on Hugging Face as promised: 2.8 trillion parameters, 104B of them active per token, a 1-million-token context, and native text, image and video input. It comes under a custom Kimi K3 License rather than the modified MIT licence of earlier Kimi releases, so users should check its terms before commercial use.

• 896 experts, 16 routed plus 2 shared per token; 93 layers mixing Kimi Delta Attention with gated attention
• Weights are stored in 4-bit MXFP4, trained to tolerate that compression, which cuts serving memory
• Model-card scores: GPQA Diamond 93.5, Terminal-Bench 2.1 88.3, BrowseComp 91.2, Video-MME 90.0