Kimi K2.6 agent swarm

Moonshot AI has released Kimi K2.6, a system built to split one big job across hundreds of AI workers running in parallel. (kimi.com) In plain terms, an “agent swarm” works like a team with many specialists instead of one chatbot taking turns on every subtask. Moonshot said K2.6 can coordinate about 300 sub-agents and handle runs with more than 4,000 tool calls. (kimi.com 1) (kimi.com 2) Moonshot said K2.6 is available on Kimi.com, in the Kimi app, through its application programming interface, and in Kimi Code. The company’s product site describes K2.6 as a multimodal model, meaning it can work with more than just text. (kimi.com) (moonshot.ai) The release lands as AI labs are shifting from single-response chatbots toward systems that can keep working for hours, call tools, and revise their own output. Moonshot’s K2.6 post highlights “long-horizon execution,” its term for jobs that unfold over many steps instead of one answer. (kimi.com) Moonshot’s examples are aimed at software work, where long runs are easiest to measure. In one company demo, K2.6 ran for more than 12 hours and over 4,000 tool calls while improving local model inference speed from about 15 to about 193 tokens per second. (kimi.com) In another example, Moonshot said K2.6 spent 13 hours modifying an open-source financial matching engine, tried 12 optimization strategies, made more than 1,000 tool calls, and changed more than 4,000 lines of code. The company said that run raised medium throughput from 0.43 to 1.24 million transactions per second and peak throughput from 1.23 to 2.86 million transactions per second. (kimi.com) Moonshot has been building toward this model for months. Its February post on Agent Swarm described a system that could spawn up to 100 sub-agents for research and synthesis; K2.6 pushes that ceiling to roughly 300. (kimi.com 1) (kimi.com 2) The company is also framing K2.6 as open-source or open-weight infrastructure rather than only a hosted assistant. Moonshot’s site says its research team shares work with the open-source community, and its K2.6 blog says the model is being open-sourced across its main products. (moonshot.ai) (kimi.com) The near-term test is whether developers trust these swarms on real work instead of demos. Moonshot’s launch pitch is that K2.6 can stay on task long enough to write, code, analyze, and restructure large projects without handing control back after every step. (kimi.com)