Agent Swarm lets one lead agent share a big job between many helpers. In April 2026 the Kimi K2.6 model made that team three times bigger: up to 300 agents and 4,000 steps in one task. Here is what changed, what it made possible, and what you use today.
From 100 agents with Kimi K2.5 to 300 with Kimi K2.6.
300 agents and 4,000 steps, up from 100 and 1,500 in Kimi K2.5.
Different kinds of agents in one job, and several kinds of file out of one run.
K3 Swarm keeps the same limits, with a stronger model in charge.
Agent Swarm first appeared as a research preview with Kimi K2.5 in January 2026. Three months later, Kimi K2.6 raised its limits. Moonshot called it "a qualitative leap".
More agents working side by side means a wide job finishes sooner. More steps means the swarm can stay on a long job without running out of road.
Figures from Moonshot's Kimi K2.6 announcement. They are upper limits, not what every task uses.
| Feature | K2.5 swarm | K2.6 swarm | K3 Swarm |
|---|---|---|---|
| Arrived | 27 January 2026 | 20 April 2026 | With Kimi K3, July 2026 |
| Agents at once, at most | 100 | 300 | 300 |
| Steps in one task | 1,500 | 4,000 | More than 4,000 tool calls |
| Status | Research preview | Built on the preview | Beta, in the Kimi apps |
| Mixed agent types in one job | Not stated | Yes | Yes |
| Several file types from one run | Not stated | Yes | Yes |
| Model's memory (context) | Not stated | About 256,000 tokens | About 1 million tokens |
| Where you find it now | Model retired from the API on 31 August 2026 | Replaced in the apps by K3 Swarm | kimi.com and the Kimi app, paid plans |
On a small screen, swipe the table sideways to see every column.
K2.5 and K2.6 details are from Moonshot's K2.6 announcement. K3 Swarm details are from Kimi's help centre. "Not stated" means the sources we read did not say.
The numbers got the headlines. These four changes matter just as much in daily use.
The swarm creates agents for different kinds of work in the same job, such as wide searching next to deep research.
A single run can produce documents, websites, slides and spreadsheets together.
Give it a PDF, spreadsheet, slide deck or Word file, and it learns the structure and style to reuse later.
A research preview where agents on different devices and models, and people, share one workspace.
If your company report always has the same sections and tone, show Kimi one good example. Later reports can follow the same pattern without you describing it again.
Kimi acts as the coordinator. It matches each task to the agent with the right skills and tools, and hands the work to another agent if one stalls or fails.
Moonshot says one of its own teams ran a K2.6 agent for 5 days without a break, watching systems and responding to problems.
These are examples Moonshot showed at the K2.6 launch, as reported at the time. They show the kind of job a swarm is good at.
100 agents each matched one CV against one of 100 job listings in California and wrote a tailored version for it. One job, 100 parts, all at once.
A research run turned an astrophysics paper into a reusable skill, then produced a 40-page report, a 20,000-entry dataset and 14 charts.
Moonshot says its own marketing team used Claw Groups to produce content and run campaigns, with K2.6 handing out the work.
These are the company's own demonstrations, not independent tests. Your results will depend on the task, your plan and your credits.
kimi-k2.6, at $0.95 in and $4 out per million tokensReady to try it? The Agent Swarm guide has the steps, plan limits and ways to save credits. New to agents? Start with Agent Mode.
Replace anything in [square brackets] with your own details. Test with a small number first.
Here is my CV and a list of [number] job adverts. For each advert, write a tailored version of my CV that keeps every fact true and brings forward the most relevant experience. Return one file per job.
Research the [industry] market in [country]. Give me three things from the same findings: a written report, a spreadsheet of the key figures with sources, and a 10-slide summary.
Study the attached report. Learn its structure, headings and tone. Then write a new report on [topic] that follows the same format exactly.
Find the [number] most cited papers on [topic] since [year]. Summarize each in three lines, then write one deep section on the three you judge most important and why.
Read all the attached documents. Write one long review that groups them by theme and compares what they say, with a table listing every document and its main point.
From these [number] web pages: [links], pull out every [type of item] with its [fields]. Build one clean dataset, remove repeats, and make three charts of the main patterns.
K2.6 is a general-purpose model built for long agent tasks and coding. At launch it was offered on Kimi's website, in the Kimi app, on the API and in Kimi Code. Guides: Kimi K2.6 · Kimi K2.6 API · Kimi K2.5 · Kimi K3 · all models
300 agents is the ceiling. Most tasks use far fewer, and your plan sets its own caps.
A job that cannot be split into parts will not run faster with more agents.
More agents do more work, so swarm tasks use more credits.
With hundreds of results, open the sources for a few at random.
Claw Groups was a research preview. It may work differently now.
The numbers here come from Moonshot. We found no independent tests.