Moonshot AI has introduced Kimi K2.5 — an open multimodal language model now available to users on Kimi.com platforms, in the Kimi app, via API, and Kimi Code. The developers focus on creating a unique “swarm agent” system: the model can simultaneously coordinate up to 100 sub-agents, performing up to 1500 instrumental calls in parallel. This reduces the execution time of complex workflows by 4.5 times compared to previous single-agent models.
Moonshot AI emphasizes that Kimi K2.5 was trained on a massive dataset of 15 trillion text and visual tokens. Such extensive training has provided the model with high performance in document creation, spreadsheet modeling, and comprehensive office task handling. The model shows consistent efficiency growth in many areas — from multi-language code generation to visual debugging.
Particularly noteworthy is Kimi K2.5’s ability to transform video recordings of website interfaces into full-fledged code while preserving all design elements and interactions. To do this, simply provide a screencast and give a simple instruction. This feature has already sparked lively discussions among developers and users testing the model’s new capabilities.
The developers have implemented the Parallel-Agent Reinforcement Learning method, which allows dynamic creation and coordination of sub-agents for distributed task execution. Kimi K2.5 is already being integrated into popular development environments, and the “swarm agent” feature is available in beta for users with high subscription levels, with gradual access expansion via API.
