Kimi's Agent Swarm Scales to 300 Sub-Agents, and Moonshot Published the Weights

Moonshot AI has pushed the most aggressive version yet of an idea the rest of the industry is still approaching carefully: instead of making one model reason longer, hand the job to a crowd of models working at the same time.

The company's Kimi K2.6, released in April 2026, can decompose a task and dispatch up to 300 sub-agents across 4,000 coordinated steps. That is a threefold jump from K2.5, which ran up to 100 sub-agents across 1,500 tool calls. Moonshot describes the arrangement in unusually plain terms on its own blog: "you are hiring a CEO, who helps find researchers, analysts, fact-checkers, all hired on the spot on its own."

What is actually new

Multi-agent orchestration is not new. Frameworks for it have existed for years. Two things separate this release from the field.

The first is that the orchestration is trained, not scripted. Moonshot says K2.5 learned to direct its own swarm through Parallel-Agent Reinforcement Learning (PARL), without predefined roles or hand-written workflows. Most competing systems require a developer to specify in advance which agent does what. Here the model decides.

The second is licensing. K2.6 ships under a Modified MIT License with weights published on Hugging Face. A capability of this kind normally arrives as a hosted API you rent. This one can be downloaded and run.

The numbers, and who is claiming them

On Moonshot's own reporting, K2.5 completed complex research and coding work between 3x and 4.5x faster than running the same steps sequentially, and reached 50.2% on Humanity's Last Exam at roughly 76% lower cost than Claude Opus 4.5. On published benchmarks, K2.6 scores 86.3 on BrowseComp against K2.5's 78.4, and 92.5 on DeepSearchQA against 78.6 for GPT-5.4.

These are vendor and vendor-adjacent figures. They have not been independently reproduced, and Moonshot itself labels the swarm "an early research preview," listing direct sub-agent communication and dynamic control of parallel width as work still to come.

Why long documents are the use case people notice first

K2.6 carries a 256K-token context window, roughly 150 to 200 pages of ordinary text. That is the practical reason the tool is spreading by word of mouth among translators and researchers rather than developers: a long report or a book chapter can go in whole, instead of being cut into fragments that lose their cross-references.

Arabic-language users have been circulating a specific claim, that a 200-page illustrated book was translated in a single pass after five other tools failed. We could not verify that account. It is directionally consistent with a 256K window, though a 200-page book dense with images sits at the very top of that range, and image-heavy documents consume context far faster than plain text.

Two different things share this name

This is worth stating plainly, because the confusion is now circulating widely. There are two unrelated projects called Agent Swarm.

One is the mode inside Moonshot's Kimi, described above, built in China, open-weight, and the source of the 300 sub-agent figure.

The other is desplega-ai/agent-swarm, an MIT-licensed open-source framework with around 670 GitHub stars, maintained by desplega.sh. It runs a lead agent that routes work to workers such as Claude Code or Codex inside isolated Docker containers, takes tasks from Slack, GitHub, email or an API, and keeps a memory of past sessions. It is a genuinely useful project. It has no connection to Moonshot or Kimi, and it does not do document translation.

Posts that attach the "300 employees" figure to the GitHub repository are merging the two. Readers looking for the book-translation behaviour want Kimi. Readers who want a self-hosted agent team wiring into their own repositories want the framework.

The part the announcements leave out

Running agents in parallel does not divide the work, it multiplies the bill. Independent research on multi-agent systems has been consistent on this point: a document-analysis workflow that costs 10,000 tokens with a single agent has been measured at 35,000 tokens across a four-agent implementation, a 3.5x multiplier before retries or error handling. Reported figures for real user requests range from 10x to 50x the tokens of a single-agent equivalent.

The failure modes are also well catalogued. Agents duplicate each other's work, loop in redundant conversation, and can over-coordinate, validating one another's output without improving the answer. Moonshot's stated design response is "productive disagreement," letting independent agents disagree to avoid groupthink structurally. Whether that holds at 300 agents rather than 100 is exactly the kind of question independent testing has not yet answered.

What to watch

The open weights are the most consequential detail. A trained orchestrator that anyone can download changes who gets to experiment with swarm architectures, and it puts a Chinese lab in front on a capability that Western labs have mostly kept behind APIs. The benchmark leads are real but self-reported. The cost profile is the open question, and it is the one that will decide whether swarms become normal infrastructure or stay an expensive way to do things a single long-running agent could have done.