Kimi Linear 48B A3B Instruct

About the Kimi family

Moonshot AI's Kimi models: K2 (open ~1T-param MoE, ~32B active), K2 Thinking, K2.5/K2.6/K2.7 Code and K3 with effort tiers, plus Kimi Linear's hybrid attention research. Known for agentic tool use and the Muon optimizer.

How this one is configured

This configuration mixture-of-experts with ~48B total / ~3B active parameters per token.

Who makes it

More from Kimi

• Kimi K2
• Kimi K2 0905
• Kimi K2 Thinking
• Kimi K2.5 (Non-reasoning)
• Kimi K2.5 (Reasoning)
• Kimi K2.6
• Kimi K2.6 (Non-reasoning)
• Kimi K2.7 Code
• Kimi K3