--- title: "Moonshot's Kimi K2.5 deploys 100 sub-agents on one task — the agent-swarm bet" date: 2026-10-19 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-kimi-k25-agent-swarm language: en --- # Moonshot's Kimi K2.5 deploys 100 sub-agents on one task — the agent-swarm bet > Moonshot AI's Kimi K2.5 (released Jan 27 2026) pairs native multimodal understanding with an "agent swarm" (智能体集群) that spins up 100 sub-agents for a single job — a bet that coordination, not just scale, is where open models win. One model answering your question is old news. One model hiring a hundred smaller versions of itself to answer your question — and finishing 4.5 times faster than doing it alone — is the bet Moonshot AI is making with Kimi K2.5. ## A multimodal model that sees and builds Kimi K2.5 was released on January 27, 2026, and Moonshot described it at launch as the **most powerful open-source model** it had shipped to date. It builds on the earlier Kimi K2 with continued pretraining over roughly **15 trillion (15T) mixed visual and text tokens**, making it natively multimodal — it reads images and video, not just text. Two capabilities stand out: - **Coding with vision.** K2.5 can turn a screenshot or a screen recording into working front-end code, including interactive layouts and scroll-triggered animations. The lab calls this "visual as code (视觉即代码)" — you show the model a design and it builds it. - **Office fluency.** It handles Word, Excel, PowerPoint, and PDF tasks end-to-end: turning a 30,000-word paper into a slide deck, or building a financial model that used to take a human analyst days, in well under an hour. ## The agent swarm The headline feature is the **agent swarm (智能体集群)**. Given a complex job, K2.5 can spin up **up to 100 specialized sub-agents** working in parallel, executing **more than 1,500 tool calls**, and — by Moonshot's measurement — completing the work **up to 4.5× faster** than a single sequential agent. The model allocates the sub-agents and breaks down the task itself; you do not pre-build the workflow. The interesting part is structural, not just speed. Parallel sub-agents can arrive at *different* conclusions and then be forced to reconcile, which the lab argues reduces groupthink on open-ended problems — useful for market research, multi-angle analysis, or "review my product plan from five expert viewpoints" prompts. K2.5 ships in four modes — **Instant, Thinking, Agent, and Agent Swarm (Beta)** — and is available through kimi.com, the Kimi App, the developer API, and **Kimi Code**, a coding tool that plugs into VSCode and Cursor. It is released as open weights (a modified MIT licence) and is built on a mixture-of-experts (混合专家, MoE) design of around **1 trillion parameters with 32 billion activated**. Moonshot says its running cost is a fraction of closed rivals like GPT-5.2-xhigh. ## Why this is a China-story Kimi K2.5 sits inside a clear trend among Chinese large models (大模型): rather than only chasing the biggest single model, labs are exploring how smaller agents coordinate. The same instinct shows up in Tencent's agent-focused Hy3 and Zhipu's tool-using GLM-4.6. Open weights matter here because they let developers self-host the orchestration layer instead of renting it. ## Running it yourself Because K2.5 is open-weight, the orchestration layer does not have to live on Moonshot's servers. A team can host the base model and wire its own agent loop, swapping in proprietary tools or private data without sending everything to a third party. That self-hosting option is precisely why the agent-swarm idea matters beyond a single product: the pattern — one planner model spawning many worker agents — can be rebuilt on top of almost any capable open model, not just Kimi. The constraint is no longer the model; it is the quality of the workflow you wrap around it. ## Trend reading According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the agent-swarm pattern shows Chinese labs are betting that intelligence scales through coordination of many small agents, not just through bigger single models. ## Honest limitations - **Agent Swarm is beta.** Behavior, cost, and reliability are still stabilizing, and the 4.5× speed claim is Moonshot's own, not independently benchmarked across all task types. - **The open-weight licence is not unrestricted.** K2.5 uses a modified MIT licence that, per published terms, requires larger companies (above certain revenue or user thresholds) to negotiate with Moonshot before offering it as a service. - **K2.5 is already a generation behind inside its own family.** Moonshot shipped K2.6 and K2.7 Code afterward, so treat K2.5 as a strong baseline rather than the current peak. - **Office automation still errs.** On genuinely complex documents, the model can misformat or misread intent; human review remains essential. ## What readers can do now - **Try Agent Swarm on a parallel-research task.** Give kimi.com's Agent Swarm mode a broad scrape-and-summarize job (e.g., "find the top creators in 100 niches") and watch how the sub-agents divide the work. - **Use Kimi Code for vision-to-frontend.** Feed it a UI screenshot or recording and compare the generated code against a manual build — this is where the multimodal edge is most visible. - **Read the licence before commercial use.** If you plan to offer K2.5 as a service, check the modified-MIT revenue thresholds so you are not surprised by a negotiation requirement later. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-kimi-k25-agent-swarm Free to quote with attribution and a link to the original.