Last updated: October 3, 2026
Kimi Agent Swarm is Moonshot AI's mode for one Kimi chat to split a job across many sub-agents and run them at the same time.
Moonshot shipped it on January 27, 2026, with Kimi K2.5. FOMA's issue that week called it not a gimmick. The lab had trained an orchestrator to divide the work. Moonshot's help center now calls the mode K3 Swarm. The posts say 300 sub-agents. The plan you can buy runs 2 to 8.
What is Kimi Agent Swarm?
A mode inside Kimi. On the web, open kimi.com/agent-swarm. In the Kimi app, pick K3 Swarm in the model switch above the chat box.
A single agent walks a task one step at a time. Agent Swarm is for a wide job, and you do not draw the org chart. The model invents the sub-agents. Moonshot's launch write-up uses three examples: the top creators in 100 YouTube niches, returned as a spreadsheet; 200-plus Paul Graham essays, filed into topic folders; and 40 PDFs turned into a 100-page literature review with citations.
How does Kimi Agent Swarm work?
One orchestrator splits the task across sub-agents created for that run. The sub-agents are frozen. Moonshot trains the coach, not the players, so a win or a mess belongs to the scheduling.
The method is PARL (Parallel-Agent Reinforcement Learning), from the K2.5 tech post. The reward checks three things: did the answer succeed, did the orchestrator actually spawn parallel workers, and did those workers finish real subtasks. The second term fights serial collapse (the coach hands the whole job to one agent). The third fights fake parallelism (a crowd of empty tasks so the chart looks busy).
Speed is scored in critical steps: the orchestrator's own overhead, plus whichever sub-agent is slowest in that stage. Extra agents only help if they shorten that path. The help center still says this cuts critical steps by 3x to 4.5x on wide search, up to 4.5x less wall-clock time than one agent in a line. That 4.5x number is the January figure. The K3 post does not publish a new one.
Sub-agents keep their own notes and report a conclusion up. Moonshot calls that context sharding. The help center says the main agent oversees about 15 steps, the same cap as the K2.5 BrowseComp swarm eval, so it may be an eval limit repeated as product copy. On that benchmark the help center reports accuracy moving from 15.9% with one agent to 33.3% with the swarm, and about 40% fewer critical steps than a "drop the context" baseline. It does not say which model produced those scores.
Which Kimi model runs Agent Swarm?
Kimi K2.5 (January 27, 2026) is the launch. Research preview, beta on Kimi, up to 100 sub-agents and 1,500 tool calls, no predefined roles. The K2.5 post said high-tier paid users got free credits to try it. That was the launch offer, not the current pricing page.
Kimi K2.6 (April 20, 2026) raised the ceiling. The K2.6 post says 300 sub-agents and 4,000 coordinated steps, still at 4.5x, with mixed jobs in one run (search plus writing, slides and spreadsheets together). Not your plan.
Kimi K3 is the current product label. The help center says Agent Swarm now runs on K3, and the app switch is named K3 Swarm. Kimi K3 itself is a separate launch: a 2.8 trillion parameter model with a 1 million token context. FOMA covered it the week of July 22, 2026. The K3 post's benchmark note is dated July 16, and it says weights would land by July 27. That post does not re-announce Agent Swarm. The multi-agent example in it is inside Kimi Work: one gravitational-wave analysis used 20+ concurrent subagents. "The swarm moved to K3" is the help center's label, not a new 300-agent benchmark.
K3's API price is $0.30 per million tokens on a cache hit, $3 on a miss, and $15 out. That bill is for the model. Swarm, as documented, is the chat mode. Moonshot has not published a swarm endpoint.
How do you use Kimi Agent Swarm?
The model picks the workers.
Open kimi.com/agent-swarm, or select K3 Swarm in the app.
Describe a wide task in one message.
Watch sub-agents, tool calls, and sites while it runs.
Download the files: code, tables, charts, or Office docs.
Later turns are supposed to route themselves between chat and agent.
How much does Kimi Agent Swarm cost?
Swarm is a paid membership feature. The help center says a run costs "roughly several times" a regular agent task, depending on how many sub-agents it spawns. Agent, Kimi Code, Kimi Work, Kimi Claw, and K3 share one credit pool, and unused credits expire at the end of the cycle.
International prices, from Kimi's USD page. Monthly first, annual in parentheses:
Moderato, $19/month ($15): Swarm on, 2 concurrent subtasks, about 60 agent credits.
Allegretto, $39/month ($31): 4 concurrent subtasks, about 150 credits. First tier with Kimi Claw.
Allegro, $99/month ($79): 8 concurrent subtasks, about 360 credits, plus K3's 1 million token chat.
Vivace, $199/month ($159): also 8 concurrent subtasks, about 720 credits, plus the long context.
The help center still says the system can direct up to 300 sub-agents. The plan you can buy caps concurrent swarm subtasks at 2, 4, or 8. Moonshot has not explained whether 300 is a ceiling inside one run or a lab number your account never schedules. The pricing page promises the small number.
China's membership page is a different ladder, in yuan, with no Vivace: Andante ¥49 (no Swarm), Moderato ¥99 (about 25 Swarm uses, 2 at once), Allegretto ¥199 (about 50 uses, 4 at once, plus Claw), Allegro ¥699 (about 120 uses, 8 at once). If a screenshot disagrees, check which domain you are on.
What is Kimi Claw vs Agent Swarm?
Different products, same credit pool.
Kimi Claw is Moonshot's one-click OpenClaw: an assistant with a personality and long-term memory. Moonshot deploys it to a cloud sandbox, you link an OpenClaw you already run, or you put it on a desktop or an Android phone. It connects to chat channels (WeChat, Feishu, WeCom, and others). Cloud deploy starts on Allegretto. The host is not free while idle: the help center says about 0.6% of your membership credits comes off every day for the sandbox, settled at 4 p.m., until you delete it. The default model on that setup is Kimi K2.6. A local OpenClaw can be pointed at Kimi K3 (kimi-coding/k3, 1 million token context). That is a config change, not the Swarm button.
Agent Swarm is a task. You hand it a wide job, it fans out, it returns files, it stops. It starts on the cheaper Moderato tier: 2 concurrent subtasks, not 300. It does not live in your messaging apps.
Claw Groups, in the K2.6 post, is a research preview: several agents and humans in one workspace, with K2.6 assigning work. Moonshot calls it a new instantiation of the swarm architecture. It is not a rename of the chat mode.
Kimi Work is the desktop app. FOMA's June 11 issue covered it as up to 300 parallel sub-agents on your machine. The K3 post still points desktop work at Kimi Work 3.1.0 or later.
Is Kimi Agent Swarm a gimmick?
The design, no. One trained orchestrator, frozen workers, and a score based on the slowest agent is a real parallel system. The January call still holds. Later posts do not walk that back.
The headcount, yes. Launch demos say 100 agents. The K2.6 post says 300. The plan you pay for runs 2 to 8 at once, and each run costs several times a normal agent task. The help page has also not replaced January's 4.5x with a K3 measurement.
Wide search, a batch download, or a long document from many sources is the job. Kimi Code is the repo tool. Kimi Claw is the assistant that keeps memory in a chat app.
What else should you compare it to?
Anthropic's multiagent orchestration, from Code with Claude in May 2026, covered by FOMA. A lead agent delegates to specialist sub-agents in parallel on a shared filesystem, each with its own model, prompt, and tools. A grader checks a rubric you wrote and sends the work back. Anthropic's recap says that was worth up to 10 points on their hardest internal tasks. You name the specialists. Kimi's pitch is that the orchestrator names them.
Hermes Agent is the other name in this search. Nous Research's local agent, covered by FOMA that May, uses a Kanban board so named agents share a task list, each worker its own process. The docs pitch that board against "fragile in-process subagent swarms." ChatGPT's coworker named Hermes is a third product.
FOMA followed this from the K2.5 launch through K3. What is FOMA.