Kimi K2.5 goes open-source with 100-agent swarms

Moonshot AI just open-sourced Kimi K2.5, a native multimodal visual agent that handles text, images, and video while orchestrating swarms of parallel sub-agents to tackle complex workflows.

Rather than tacking a vision module onto a text model, Moonshot trained K2.5 on 15 trillion mixed visual and text tokens. The result is an autonomous system capable of deep reasoning across media types and real-world tools.

What it can do

The highlight here is Agent Swarm, currently in beta. Instead of running a single agent sequentially through a long checklist, Kimi K2.5 breaks a large goal into parallel tracks.

It can coordinate up to 100 sub-agents working simultaneously, handling up to 1,500 tool calls along the way. Moonshot reports that this multi-agent orchestration speeds up task completion by 4.5x compared to single-agent setups.

It also leans heavily into visual coding. Users can feed it a design image or a screen-recorded video, and the model turns the visual input into a responsive website complete with interactive motion.

Benchmark gains and developer tooling

On the leaderboard, Kimi K2.5 sets new highs for open models and beats several frontier closed models. It achieved 50.2% on the full HLE benchmark and 74.9% on BrowseComp.

For visual tasks, it hit 78.5% on MMMU Pro and 86.6% on VideoMMMU, topping GPT 5.2 and Claude Opus 4.5 in video reasoning. In software engineering, it scored 76.8% on SWE-bench Verified and outperformed Gemini 3 Pro and GPT 5.2 on multilingual coding.

Alongside the model, Moonshot released Kimi Code, an open-source terminal coding assistant aimed squarely at Anthropic's Claude Code and Google's Gemini CLI. It hooks into editors like VSCode, Cursor, and Zed, accepting both images and videos as direct inputs.

How to get started

The model weights and inference code are openly available on Hugging Face. Developers looking to build on it programmatically can access the Moonshot API.

If you want to test it without hardware overhead, Kimi K2.5 is live on kimi.com in chat and agent modes, while the Agent Swarm feature is in beta for high-tier accounts. Kimi Code is available via its dedicated portal.

Why it matters

If you build software or design interfaces, Kimi K2.5 is worth immediate testing. Turning raw UI screen recordings directly into functional, animated code drastically lowers the barrier to prototyping.

More importantly, it provides an open alternative to proprietary developer tools like Claude Code. Proving that parallel sub-agent swarms can run reliably on an open-weights model puts immense pressure on closed-source labs to open up their own agent orchestration layers.