Moonshot has released Kimi K2.6, an open-source native multimodal agentic model designed to advance practical capabilities in long-horizon coding and autonomous task orchestration.

  • The model supports long-horizon coding across languages like Rust, Go, and Python, as well as coding-driven design for generating production-ready interfaces.
  • It features an elevated agent swarm capable of scaling horizontally to 300 sub-agents executing 4,000 coordinated steps for parallel task decomposition.
  • Kimi K2.6 demonstrates proactive orchestration for persistent background agents managing schedules and cross-platform operations without human oversight.
  • In evaluations on Humanity's Last Exam (HLE), the model achieved 36.4% accuracy without tools and 55.5% with tools, using a context length of 262,144 tokens.

The release provides libraries, inference providers, and local app settings to facilitate integration with Transformers, vLLM, SGLang, and Docker.