Kimi K3 Open Weights Land on Hugging Face: 2.8T Total / 104B Active, 1M Context, and a New License With a $20M Revenue Trigger

simonwillison.net (corroborated by the moonshotai/Kimi-K3 model card)·high signal

Moonshot released the Kimi K3 weights on July 27 — a 1.56TB Hugging Face repo of a 2.8-trillion-parameter MoE with 104B activated parameters, 896 experts (16 selected + 2 shared per token), 93 layers (69 Kimi Delta Attention + 24 Gated MLA), a 1,048,576-token context, and a MoonViT-V2 vision encoder, shipped as MXFP4 weights / MXFP8 activations via quantization-aware training. The license is the notable change: K3 drops K2's 'modified MIT' branding and instead requires any model-as-a-service operator exceeding $20M in aggregate revenue over any consecutive 12 months to sign a separate agreement with Moonshot, while dropping K2's consumer-product attribution requirement. Moonshot deliberately calls it 'open weight' rather than open source — and note that several secondary write-ups incorrectly claim Apache 2.0, which the model card contradicts.