Kimi

An AI assistant and open-weight model family built by Moonshot AI, known for long-context chat and agentic tool use (including Kimi K2 / K2.6).

How it works

Kimi models use the MoE Transformer architecture with MLA for memory-efficient long-context attention. Kimi K2 was trained with the Muon optimiser and QK-Clip for training stability. The model family is designed for agentic use: Kimi K2.6 features agent swarm capabilities where it can orchestrate thousands of parallel sub-agents for complex tasks. Kimi is also notable for its long-context ability, handling 1M+ token contexts via techniques like Ring Attention.

Why it matters

Kimi K2 demonstrated that a Chinese open-source MoE model could match GPT-4-class performance on agentic and coding tasks, contributing to the growing evidence that frontier AI capabilities are no longer the exclusive domain of US labs. Its open-weight release enabled the research community to study and extend MoE architecture at scale. Kimi K2.6's agent swarm approach represents a leading example of how agentic AI capabilities are being pushed forward in practice.