Moonshot releases Kimi K2.5 with agent swarms

Moonshot AI introduced Kimi K2.5 in a technical blog post, calling it "the most powerful open-source model to date." The model builds on Kimi K2 with continued pretraining on about 15T mixed visual and text tokens. Moonshot says it is state of the art among open models for coding, especially front-end work, and for vision.

The company highlights "coding with vision": generating code from images and video, and visual debugging, where the model looks at its own rendered output and revises it. Moonshot argues that at scale, joint vision-text pretraining removes the trade-off between the two modalities. Its coding claims rest partly on Kimi Code Bench, an internal benchmark. The post also cites results on HLE, BrowseComp and SWE-Verified at "a fraction of the cost," but the excerpt does not give the scores.

The main technical contribution is Agent Swarm, released as a research preview. K2.5 can spin up to 100 sub-agents running in parallel across up to 1,500 tool calls, with no predefined roles or workflows. Moonshot says this cuts execution time by up to 4.5x compared with a single agent. The swarm is trained with Parallel-Agent Reinforcement Learning (PARL): a trainable orchestrator splits a task into subtasks and hands them to frozen sub-agents. The reward combines task outcome with two auxiliary terms. One rewards launching sub-agents, to stop the orchestrator from falling back to working alone. The other rewards completed subtasks, so it can't game the parallelism reward by spawning agents that do no useful work. Both auxiliary weights are annealed to zero during training.

K2.5 is available on Kimi.com, the Kimi App, the API and the open-source Kimi Code terminal tool. Agent Swarm is in beta, with free credits for high-tier paid users.