Moonshot Releases Kimi K3 Weights With Revenue-Triggered License

Moonshot released Kimi K3's weights on Hugging Face on July 27, a 1.56TB repository holding a 2.8 trillion parameter mixture-of-experts model with 104 billion active parameters. The license is the change: any operator clearing $20 million in revenue over any rolling 12 months must sign a separate agreement with Moonshot. K2 never carried that term.

The architecture: 896 experts, 16 selected plus 2 shared per token, across 93 layers split between 69 Kimi Delta Attention and 24 Gated MLA. It runs a 1,048,576-token context window and a MoonViT-V2 vision encoder. Moonshot trained it with quantization-aware training and shipped it as MXFP4 weights with MXFP8 activations, per the model card.

Moonshot calls this open weight, not open source. K2 shipped under a modified MIT label; K3 drops that framing, and also drops K2's requirement that consumer products credit Kimi in their branding. Per Simon Willison's write-up, several secondary reports already got the license wrong, claiming Apache 2.0 terms the model card contradicts.

What this changes

Small teams, researchers, and hobbyists get an unencumbered model. Anyone who builds a real business on it owes Moonshot a conversation, narrower than a permissive license and more enforceable than an attribution clause. Watch whether other labs releasing frontier-scale weights start writing revenue thresholds into their own terms instead of relying on credit requirements.