Kimi K2.6 is released and open-sourced, with the code and Agent cluster capabilities greatly improved.

On the dark side of the moon, Kimi has released and open-sourced the Kimi K2.6 model, significantly improving its comprehensive abilities such as general agent, code, and visual understanding. It outperforms or matches closed-source models such as GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro in multiple benchmark tests. The long-range encoding capability of Kimi K2.6 has been significantly enhanced, allowing it to continuously encode for 13 hours in tests and write or modify over 4000 lines of code. At the same time, the autonomous execution capability of the Agent has been greatly strengthened. The Agent cluster architecture driven by the K2.6 model has been greatly upgraded, supporting 300 sub-Agents to perform 4000 collaborative steps in parallel, achieving greater scalability. In response to high-load workflows and active agent frameworks such as OpenClaw and Hermes Agent, K2.6 has the capability of automated task processing, supporting continuous autonomous operation for up to 5 days.