Moonshot AI releases Kimi K2, an open-weight 1-trillion-parameter MoE model

Moonshot AI/Kimi models

Moonshot released Kimi K2 under a modified MIT licence: a mixture-of-experts model with 1 trillion total parameters (32B active per token), trained on 15.5 trillion tokens with the Muon optimizer without training instabilities. The instruct version scored 65.8% on SWE-bench Verified, strong for an open model built for agent tasks.

• 1 trillion total parameters, 32B active per token; trained on 15.5T tokens with Muon optimizer.
• Zero training instabilities; novel optimization techniques for stability at scale.
• Kimi K2-Instruct: 65.8% on SWE-bench Verified; 69.6% on AIME 2024; 97.4% on MATH-500.
• Available under modified MIT licence; includes chat-optimized Instruct variant with tool-calling.