Moonshot AI released Kimi K2.6 on Monday, pitching its open-weight model as a match for GPT-5.4 and Claude Opus 4.6 on coding benchmarks. The model weights are live on Hugging Face under a modified MIT license.
The real story is scale. K2.6 can run 300 sub-agents in parallel at 4,000 steps each, up from 100 and 1,500 in K2.5. Moonshot's tech blog says the model handles 12-plus hours of continuous execution across Rust, Go, and Python, with over 4,000 tool calls per run.
The benchmark table is self-reported, and independent testing hasn't confirmed the gains. K2.6 posts 54.0 on HLE with tools, 58.6 on SWE-Bench Pro, and 83.2 on BrowseComp, edging Claude Opus 4.6 on the first two per Moonshot's own chart. Pure reasoning is where the model still trails closed rivals.
Access runs through kimi.com chat and agent modes, the Moonshot API, and Kimi Code for production work. A research preview called Claw Groups, which mixes user-owned bots with human collaborators, shipped alongside.
Bottom Line
K2.6 triples K2.5's parallel agent count to 300 and extends runs to 4,000 steps each.
Quick Facts
- SWE-Bench Pro: 58.6 (company-reported)
- 300 parallel sub-agents, 4,000 steps per run
- 12+ hours continuous execution across Rust, Go, Python
- Modified MIT license; 100M+ MAU or $20M+ monthly revenue deployments must credit Kimi K2.6 in UI
- HLE with tools: 54.0; BrowseComp: 83.2 (self-reported)
