Kimi K2.6 ships with 300-agent swarms
Moonshot just dropped Kimi K2.6, and the headline number is wild. The model can dynamically scale to 300 sub-agents executing 4,000 coordinated steps in a single task. The previous K2.5 capped at 100. Multi-agent went from a research toy to a production primitive in one release.
The coding numbers back it up. 80.2% on SWE-Bench Verified, 76.7% on SWE-Bench Multilingual, 89.6% on LiveCodeBench, 66.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro. That puts K2.6 next to Claude Opus 4.7 and GPT-5.2-Codex on most coding leaderboards, but with open weights and access through kimi.com, the Kimi App, the API, and Kimi Code.
Moonshot is also leaning hard into a feature called Claw Groups, basically multi-agent collaboration where specialized agents handle demos, benchmarking, social, video, all coordinated by one K2.6 instance. Their own marketing team apparently runs end-to-end content production this way. They are dogfooding the agent-swarm pitch instead of just announcing it.
The deeper story is the rate. Kimi shipped K2 a few months ago, K2.5 in March, K2.6 now. Each step adds real capability, not just bigger numbers. The Chinese open-weight stack is moving faster than most people expect, and the gap to closed labs on agentic coding is now visibly small.
Link kimi.com/blog/kimi-k2-6
← Back to all articles
The coding numbers back it up. 80.2% on SWE-Bench Verified, 76.7% on SWE-Bench Multilingual, 89.6% on LiveCodeBench, 66.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro. That puts K2.6 next to Claude Opus 4.7 and GPT-5.2-Codex on most coding leaderboards, but with open weights and access through kimi.com, the Kimi App, the API, and Kimi Code.
Moonshot is also leaning hard into a feature called Claw Groups, basically multi-agent collaboration where specialized agents handle demos, benchmarking, social, video, all coordinated by one K2.6 instance. Their own marketing team apparently runs end-to-end content production this way. They are dogfooding the agent-swarm pitch instead of just announcing it.
The deeper story is the rate. Kimi shipped K2 a few months ago, K2.5 in March, K2.6 now. Each step adds real capability, not just bigger numbers. The Chinese open-weight stack is moving faster than most people expect, and the gap to closed labs on agentic coding is now visibly small.
Link kimi.com/blog/kimi-k2-6
Comments