Moonshot updated Kimi to version 2.6

The model is now much stronger in long coding tasks and outperforms Composer 2 from Cursor (which is based on K2.5) in benchmarks. Additionally, the company continues to invest in agent ROI — K2.6 was trained to use up to 300 sub-agents in parallel, while other teams do not focus on this as much (which is a mistake). They also say that the model now works better with OpenClaw and other similar agents.