Kimi K2, An Open-weight Agentic Model From Moonshot AI

Moonshot AI's Kimi K2, a 1-trillion-parameter open-weight model for agentic AI, uses a Mixture of Experts architecture to achieve top performance in knowledge, math, and coding while reducing costs. It features Multihead Latent Attention for efficient inference, the MuonClip optimizer for stable training, and data rephrasing for token efficiency. Kimi K2 employs flexible training infrastructure for long contexts and incorporates agentic data synthesis and reinforcement learning, marking a major advance in autonomous AI systems.