Moonshot AI Launches 1T-Parameter Kimi-K2.6 Model

Moonshot AI has released Kimi-K2.6, a one-trillion-parameter open-source large language model that the Chinese startup claims outperforms GPT-5.4 and Claude Opus 4.6 on several benchmarks. The model features SwiGLU activation functions, multi-head latent attention, and a 400-million-parameter vision encoder for multimedia processing. Kimi-K2.6 uses a mixture-of-experts architecture with 384 specialists, activating only eight per query to reduce hardware demands. It can deploy up to 300 agents for complex tasks and includes a human-AI collaboration feature called claw groups. The model topped the challenging HLE-Full benchmark with a score of 54, edging out Opus 4.6 and GPT-5.4.