Moonshot releases Kimi K2.5: a swarm of 100 subagents and 1500 tool calls

Moonshot AI has introduced Kimi K2.5, a multimodal model that the company calls the strongest open model available today. It is built on Kimi K2 and further trained on roughly 15T mixed visual and text tokens. The main new feature is Agent Swarm: the model deploys up to 100 subagents on its own and makes up to 1500 tool calls. According to the company, this cuts the execution time of complex tasks by up to 4.5x compared with a single agent.

How the swarm works

Subagents are not defined by hand. The orchestrator breaks a task into parallel subtasks and creates workers on the fly, such as a researcher, a physicist, or a fact-checker. It was trained with Parallel-Agent Reinforcement Learning (PARL). The orchestrator is trainable, while the subagents are frozen. As a result, feedback is delayed, sparse, and non-stationary.

This setup has two typical failure modes. The first is serial collapse: the orchestrator slides into single-agent execution even though it could parallelize the work. The second is spurious parallelism: it spawns subagents to inflate a metric without actually decomposing the task. To counter the first, the reward includes a bonus for creating subagents; to counter the second, it includes a reward for the share of completed subtasks. Both coefficients are annealed to zero over the course of training, leaving only the reward for the final outcome.

Another design choice is the Critical Steps metric, an analogue of the critical path in parallel computing. At each stage it sums the steps of the main agent and of the slowest subagent, so extra subagents help only when they shorten the critical path.

• Scale: up to 100 subagents, up to 1500 coordinated steps.
• Speed: up to 4.5x faster execution; in internal tests, latency drops by 80%.
• Critical Steps: wide-search scenarios need 3–4.5x fewer.
• Context: 256K tokens in tests.
• HLE without tools: 31.5 (text) and 21.3 (images).
• HLE with tools: 51.8 (text) and 39.8 (images).
• Office benchmarks: +59.3% (AI Office) and +24.3% (General Agent) over K2 Thinking.

Code, vision, and office work

Moonshot emphasizes frontend work: from a single prompt, the model builds interfaces with animations, including scroll effects. It can reconstruct a website from a video and debug layouts visually, looking at its own output and iteratively fixing it. According to the authors, at this scale of pretraining the trade-off between vision and text disappears: both capabilities improve together.

For development, the company has released Kimi Code, an open-source terminal tool. It integrates with VSCode, Cursor, and Zed, accepts images and video, and finds and migrates existing skills and MCP setups on its own. The company also describes an office scenario: Word documents with annotations, financial models with pivot tables, LaTeX formulas in PDFs, and documents up to 100 pages.

The numbers come with caveats. The developer ran the tests itself. The comparison included DeepSeek-V3.2, Claude Opus 4.5, GPT-5.2 (xhigh), and Gemini 3 Pro, and in the visual tests also Qwen3-VL-235B-A22B-Thinking. GPT-5.2 xhigh was not evaluated on every test because of service instability. In the visual tests it had about 10% failures, which were counted as errors, and the authors acknowledge that its score is probably understated. Some metrics, including Kimi Code Bench and the office benchmarks, are internal. To reproduce the results, Moonshot recommends the official API or providers verified through Kimi Vendor Verifier.

Where to start

Developers can reasonably start with Kimi Code paired with K2.5: the tool is open, plugs into familiar IDEs, and you can test the visual debugging on your own frontend. Teams working on wide-search and data-collection tasks should test Agent Swarm. It is currently in beta on Kimi.ai, and users on the higher paid plans get free credits.

When comparing with competitors, measure not only accuracy but also execution time and cost, since the swarm is designed to improve exactly those. If you run the model with a third-party provider, check it against Kimi Vendor Verifier; otherwise your results may diverge noticeably from the published ones. Moonshot has only promised the full technical report, including system prompts.