Moonshot AI
Kimi models you can run locally
Moonshot AI develops Kimi, a family of long-context multimodal and agentic mixture-of-experts models.
Available models
Ordered from the lightest model to the largest.
Kimi K2.7 Code
1T · 32B active · MoE
Coding-focused agentic MoE built on K2.6 — 32B active, long-horizon coding with ~30% fewer thinking tokens
256K ctx
Kimi K2.6
1.06T · 32B active · MoE
Natively multimodal 1T MoE — 32B active, frontier agentic
256K ctx
Kimi K3
2.8T · 104B active · MoE
Frontier 2.8T multimodal MoE — 104B active, native video understanding, 1M context
1024K ctx
Kimi K3.1 Pro
2.8T · 104B active · MoE
K3 post-trained with large-scale RL — 2.8T multimodal MoE, 104B active, sparse attention for its 1M context, stronger coding and agentic work
1024K ctx