moonshotai/Kimi-K3

Summary

Moonshot AI released the weights of its massive K3 model, which has 2.8 trillion parameters and totals 1.56TB on Hugging Face.

Why it matters

K3 represents a significant leap in model scale and could advance AI capabilities. Its licensing strategy—shifting from a simple MIT modification to requiring separate agreements for high-revenue Model-as-a-Service (MaaS) operators—signals a shift in how AI companies balance openness with commercial control.

Highlights

• K3 model size: 2.8 trillion parameters, 1.56TB of weights.
• License change: No longer a 'modified MIT'; large MaaS operators must sign a separate agreement if annual revenue exceeds $20M.
• OpenRouter integration: K3 is offered through seven providers with consistent pricing ($3 million per input, $15 million per output).

Critical take

The new licensing terms could introduce commercial barriers that weaken the promise of 'open weights' and potentially limit community-driven innovation and broader accessibility.