Guides
What Is Kimi Agent Swarm? K3 Swarm Mode and Its 300 Sub-Agents, Explained
Kimi Agent Swarm (K3 Swarm) is Moonshot AI’s mode that splits one job across up to 300 parallel sub-agents. How it works, who can use it, and when it beats one agent.
Kimi Agent Swarm is a mode inside Moonshot AI’s Kimi app where one AI boss hands a big job to many AI helpers at once. People also call it Kimi swarm, K3 Swarm or swarm mode. With Kimi K3.1 Pro in the news this week, a lot of new Kimi users are spotting “K3 Swarm” in the model menu and wondering what it does.
Here is the short version: what it is, how it works, what it costs, and the kind of task where it is worth trying.
What is Kimi Agent Swarm?
Most AI agents work like one person with a to-do list. They search, read, write, then search again, one step after another. Agent Swarm works like a manager with a big team. The lead agent, which Moonshot calls the orchestrator, looks at your request, decides which parts can run at the same time, and starts sub-agents to do them. When they finish, it merges their work into one result.
The sub-agents have no fixed job titles. Moonshot’s help center says there are no predefined roles or hand-built workflows. The orchestrator invents the team it needs for each task. One run might use five helpers. Another, like Moonshot’s own example of finding the top creators in 100 YouTube niches, used 300.
Kimi swarm timeline: K2.5, K2.6 and K3 Swarm
| Version | Date | Max sub-agents | Tool calls per task |
|---|---|---|---|
| Kimi K2.5 Agent Swarm (first beta) | January 27, 2026 | Up to 100 | Up to 1,500 |
| Kimi K2.6 upgrade | April 20, 2026 | Up to 300 | 4,000+ |
| K3 Swarm (current) | Since Kimi K3 (July 2026) | Up to 300 (no new figure published) | 4,000+ (no new figure published) |
All of these numbers come from Moonshot. No outside lab has confirmed them. The help center says the swarm is “currently powered by Kimi K3”. It does not say that the new K3.1 Pro model runs it yet.
How does Kimi Agent Swarm work?
The trick is in the training. Moonshot trained the orchestrator with a method it calls Parallel-Agent Reinforcement Learning, or PARL. Only the orchestrator learns. The sub-agents are fresh copies started for each run.
Moonshot also names the ways a swarm can fail, which is useful to know before you pay for one:
- Serial collapse: the lead agent falls back to doing things one at a time, so 300 helpers act like one.
- Fake parallelism: the work looks split up, but the pieces depend on each other, so nothing gets faster.
On the BrowseComp web research test, Moonshot reports that accuracy rose from 15.9% with a single agent to 33.3% with the swarm. Again, that is the company’s own figure.
How to use Kimi Agent Swarm
- On the web: go to kimi.com/agent-swarm.
- On iPhone or Android: open the Kimi app and pick “K3 Swarm” from the model switch above the chat box.
- Then: describe the whole job in one message and send it. You can watch the task list, the sub-agents being created, the sites they visit and each agent’s status in real time.
Results can be a folder of files, a chart, a code project, or a Word, PDF, Markdown or slide document.
Kimi swarm pricing: is it free?
No. Kimi’s help center lists Agent Swarm as a beta for paid members on the Moderato, Allegretto and Allegro plans. It also warns that each swarm task uses several times the credits of a normal Agent task, depending on how many sub-agents it starts. Moonshot has not published a per-task price, so the real cost depends on your plan and how big the job is.
Kimi Agent Swarm vs a single agent vs deep research
| Kimi Agent Swarm | One agent or deep research | |
|---|---|---|
| How it works | Many sub-agents at once, merged by a lead agent | One agent, one step at a time |
| Best for | Wide jobs: 100 searches, 200 articles, 40 PDFs | Deep jobs where each step depends on the last |
| Speed | About 4.5x faster on wide jobs (Moonshot claim) | Slower on wide jobs, fine on short ones |
| Cost | Several times the credits of a normal Agent task | Lower |
A good test: could you hand the job to 20 interns who never talk to each other? If yes, a swarm fits. If each step needs the answer to the one before, a single agent is just as good and cheaper. Claude has a related idea for developers in Claude dynamic workflows, where one agent writes a program that runs many agents on a job.
What this means for you
Kimi Agent Swarm shows where AI is heading: away from one chatbot that does everything, toward teams of agents that split the work. You do not need 300 agents for most daily tasks, though. Planning a trip, writing a report or comparing three options works better with a few focused experts that read each other’s work. We explain why in why two AI agents beat one.
That is the idea behind our iPhone app. You give one brief to a small team of AI specialists, and they build on each other’s answers instead of starting from scratch.
Quick answers
What is Kimi Agent Swarm? It is a multi-agent mode in Moonshot AI’s Kimi where a lead agent splits one task across up to 300 sub-agents that work in parallel, then merges their results.
What is Kimi K3 swarm? K3 Swarm is the current name of Agent Swarm in the Kimi app’s model menu. It means the swarm now runs on Moonshot’s Kimi K3 model.
How does Kimi Agent Swarm work? An orchestrator trained with Parallel-Agent Reinforcement Learning (PARL) decides which parts of a job can run at once, starts sub-agents with no fixed roles, and combines what they find.
Is Kimi Agent Swarm free? No. It is a beta for paid Kimi plans (Moderato, Allegretto and Allegro), and each swarm task uses several times the credits of a normal Agent task.
Is Kimi swarm on GitHub or open source? The swarm mode itself runs inside Kimi’s apps and website. Moonshot publishes its help docs on GitHub and releases open model weights, but we found no official open-source release of the swarm product.
Kimi Agent Swarm vs Claude: which is better? They do different things. Kimi’s swarm is a ready-made mode for huge parallel jobs. Claude’s closest match is dynamic workflows in Claude Code, aimed at developers. For everyday tasks, a small team of agents is usually enough.
New to these words? Start with what agentic AI is or browse the agentic AI glossary.
Sources: Kimi help center: Agent Swarm · Kimi Agent overview · The Decoder: Kimi K2.5 and Agent Swarm · KDnuggets: I tried Kimi Agent
