Moonshot AI has released Kimi K3 on Hugging Face—the first open-weight model at the ~3 trillion parameter scale. The model uses a Mixture-of-Experts architecture with 896 experts (16 active per token, 104B activated parameters), supports up to 1M token context, and handles both text and images natively. It introduces Kimi Delta Attention and achieves roughly 2.5× more efficient scaling than its predecessor Kimi K2, targeting long-horizon agentic tasks like repository analysis, GPU kernel optimization, and compiler engineering.

Kimi K3 on Hugging Face