Kimi Agent Swarm Guide: How It Works, Access, Limits [2026]

A normal agent works through a task one step after another. Agent Swarm changes the shape of the work. A main agent reads the task, breaks it into parts that do not depend on each other, starts a sub-agent for each part, and then combines what they return.

Moonshot's help center describes it as a mode for large-scale search, long-form writing, and batch processing. It also says to use Swarm when you need more than one output file. A regular Agent task usually produces one file at a time, so a Word document and a slide deck in the same task is a Swarm job.

Take the research request from the original version of this post: summarize 10 competitor products and write one comparison. A single agent reads them in sequence. In Swarm, each product can go to its own sub-agent, and the main agent writes the comparison once they finish. That example is ours, not from Moonshot.

The two get confused because the product picker lists them side by side. On kimi.com the model menu lists K2.6, K3, and K3 Swarm, plus a reasoning effort setting. You can see this in the picker screenshot in our K3 review.

Moonshot's help pages use two names for the same idea. The Agent Swarm page and the plan documentation say K3 Swarm. The model-selection page says K3 Cluster and, under "Agent and Agent Cluster," says: "Agent cluster: Supports more than 4,000 parallel tool calls and can autonomously coordinate up to 300 subagents." We treat them as the same feature under two labels, but Moonshot does not say so in the pages we could check.

Downloading the K3 weights does not give you Swarm. The weights are the model. The task splitting, sub-agent scheduling, and merging are a hosted workflow, and a self-hosted setup would need its own orchestration. Our open-source model ranking covers the open-weight side.

Moonshot trained the swarm behavior with a method it calls PARL, Parallel-Agent Reinforcement Learning. The orchestrator is the part that learns. It decides how to decompose a task and when to start sub-agents. The sub-agents it launches are frozen and are created on demand, with roles such as researcher or fact-checker, rather than being a fixed set.

Moonshot measures progress by critical steps, the length of the longest chain of steps that must happen in order, instead of total steps. A swarm that launches many agents without shortening that chain gets no credit, which discourages a model that looks parallel but is not. Moonshot's 4.5x figure comes from the same measure: it reports a 3x to 4.5x reduction in critical steps in large-scale search, which translates to up to 4.5x shorter execution time.

For K3, Moonshot's model-selection page gives the same ceilings as K2.6: up to 300 sub-agents and more than 4,000 parallel tool calls. Its Agent Swarm page words the second figure as parallel workflows of up to 4,000 steps. Moonshot's help pages mix "tool calls" and "steps" without defining either, so we report both labels and do not treat them as different limits. If a task depends on the exact ceiling, test it.

Moonshot does not publish a prompt format for Swarm, so this is our advice, not documentation. Swarm pays off on wide work: many similar items, many sources, or several deliverables. Say how many items there are and what each should return. If you want a Word file and a slide deck, ask for both in the same task, since that is the case Moonshot points to for Swarm. For narrow, sequential work, choose regular K3 or Agent mode instead.

For API work, Moonshot's help pages describe Swarm as a feature of its apps. We did not find it documented as an API mode, so plan API projects around the K3 model and see the K3 review for rates and access.

Swarm is a membership feature. Moonshot lists four plans, and the cheapest one does not include it. Prices below are Moonshot's monthly prices in CNY, which we have not converted.

Credits are the real meter. Moonshot's membership page says the Agent usage figures "are estimates based on typical task token consumption, assuming your monthly credits are used on a single feature," and that credits are "calculated based on actual token consumption." So treat about 25, 50, or 120 Swarm uses as Moonshot's estimate, not a count it enforces. K3, K3 Swarm, and the other membership features draw from one shared pool, and credits come off according to actual token use. Moonshot says a Swarm task uses roughly several times the credits of a regular Agent task, and the exact amount depends on task complexity and how many sub-agents run. It does not publish a per-run credit price, so we cannot give you a cost per task. Run a small version first and watch how much of your pool it takes before you commit a large job.

• The entry plan has no Swarm. On the other plans, Moonshot estimates about 25, 50, or 120 Swarm uses, but these are estimates from typical token use. Credits are what actually come off, and a heavy task can use more than one "use" worth.
• The plan table caps parallel subtasks at 2, 4, or 8. The pages we checked do not say how that relates to the 300 sub-agent ceiling, so treat 300 as the product's upper bound, not what your plan runs at once.
• A Swarm task uses several times the credits of an Agent task, from a shared pool that also covers K3 chat and other features.
• Runs take longer than the 5 to 20 minutes typical for an Agent task, so it does not suit interactive back-and-forth.
• The 4.5x speedup is Moonshot's own figure from large-scale search. It is not an independent measurement.
• Tasks with strict step-by-step dependencies gain little from parallel agents.
• Swarm is a Kimi-hosted workflow, so your tasks run on Moonshot's service. Check your data policy before sending anything sensitive.

Use it for:

• Searching or reviewing many sources and producing one consolidated report
• Batch jobs where the same operation applies to dozens of items
• Long-form writing that can be drafted in sections and merged
• Tasks that need several deliverables from one brief

Use regular K3 or another model for:

• Quick questions and interactive chat, where K3's slower output is already a cost, as covered in the K3 review
• Sequential coding or debugging, where each step depends on the last
• API-based pipelines, since Swarm is documented as an app feature

For where Kimi sits against Claude, GPT, and Gemini on other kinds of work, see the task-by-task AI model guide. Autonomous single-agent tools such as Manus AI take a different route to long tasks.

• Moonshot AI: Kimi Agent Swarm help center page (K3 Swarm, 300 sub-agents, 4,000 steps, plan availability, credit use, how to access)
• Moonshot AI: Kimi Agent features and usage limits (credit pool, task duration, multi-file output)
• Moonshot AI: membership pricing and plan overview (plan prices in CNY, Swarm uses, parallel subtasks)
• Moonshot AI: how to choose a Kimi model and mode (K3 Cluster, 300 sub-agents, 4,000 parallel tool calls, reasoning effort)
• Moonshot AI: Kimi K2.5 tech blog (launch Agent Swarm, 100 sub-agents, 1,500 tool calls, PARL)
• Moonshot AI: Kimi K2.6 announcement (300 sub-agents, 4,000 tool calls, April 20, 2026)
• Our Kimi K3 review, for the model picker and K3 release details

What is Kimi Agent Swarm?

Agent Swarm is a mode in Kimi's products where a main agent splits a large task into parts and hands them to sub-agents that run in parallel. Moonshot's help center says the main agent can direct up to 300 sub-agent instances and run parallel workflows of up to 4,000 steps. It is built for large-scale search, long-form writing, and batch processing.

Is Agent Swarm the same as Kimi K3?

No. K3 is the model. Agent Swarm is a product mode that uses a model to coordinate many agents. Moonshot's help center says Agent Swarm is now powered by K3 and calls the mode K3 Swarm in the model picker. The released K3 weights do not include the hosted swarm workflow.

Which Kimi plans include Agent Swarm?

Moderato (CNY 99 per month), Allegretto (CNY 199), and Allegro (CNY 699). The entry Andante plan (CNY 49) does not include it. Moonshot estimates about 25, 50, and 120 Swarm uses on those plans, with 2, 4, and 8 subtasks running in parallel. The use counts are estimates based on typical token consumption, not hard quotas. Check Moonshot's membership page for current terms.

How much does Agent Swarm cost to run?

There is no separate per-run price. Swarm draws from the same credit pool as the rest of a Kimi membership, deducted by actual token use. Moonshot says a Swarm task uses roughly several times the credits of a regular Agent task, depending on complexity and the number of sub-agents.

Can I run Agent Swarm myself with the open K3 weights?

You can host the K3 model, but the weights alone do not include the hosted Swarm workflow. A self-hosted setup needs its own orchestration layer to split tasks, run sub-agents, and merge results.

Written by

Paras Tiwari

Founder, Spectrum AI Labs

Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.

More about Paras →