Moonshot AI released Kimi K2.6 on April 20, pushing its open-weight model's Agent Swarm up to 300 parallel sub-agents running 4,000 coordinated steps per task. The company detailed the expansion on its tech blog, with weights posted to Hugging Face under a Modified MIT License.
That's a jump from 100 sub-agents and 1,500 steps in K2.5, which launched in January. The underlying architecture didn't change: still a 1 trillion-parameter Mixture-of-Experts with 32B active per token and a 256K context window. Moonshot says the gains came from additional post-training focused on long-horizon stability and swarm coordination.
The pitch is that one run produces actual deliverables, not chat. Moonshot's own showcase cites a 40-page, 7,000-word research paper generated alongside a 20,000-entry dataset and 14 charts. An X post from the Kimi account goes further, claiming single runs of 100,000-word literature reviews. Independent replications haven't surfaced.
K2.6 ships in four modes: Instant, Thinking, Agent, and Agent Swarm. Moonshot's benchmark numbers put it at 58.6% on SWE-Bench Pro, narrowly ahead of GPT-5.4 and Claude Opus 4.6, though every score is self-reported. K2.6 pricing hasn't been published. K2.5 sat at $0.60 per million input tokens and $2.50 per million output.
Bottom Line
Moonshot says K2.6 coordinates 300 sub-agents across 4,000 steps in one autonomous run, but the benchmarks behind the claim are company-run.
Quick Facts
- Agent Swarm: 300 sub-agents, 4,000 coordinated steps per run
- Previous K2.5 cap: 100 sub-agents, 1,500 steps
- Release date: April 20, 2026
- Architecture: 1T-parameter MoE, 32B active, 256K context
- SWE-Bench Pro: 58.6% (company-reported)
- License: Modified MIT, weights on Hugging Face
