A new and powerful open-source model has been introduced, named Kimi K2.5. Building on its predecessor, it was trained on a massive dataset of approximately 15 trillion visual and text tokens.

The announcement highlights three core areas of advancement.

First is its strength in coding with vision. Described as the most powerful open-source model for coding to date, K2.5 can turn conversations into complete front-end interfaces and reason over images and video to generate code or debug visually. An example provided shows the model analyzing a maze image, writing code to find the shortest path, and visually overlaying the solution.

Second is a new "Agent Swarm" capability, currently in beta. For complex tasks, K2.5 can reportedly self-direct a swarm of up to 100 sub-agents to execute parallel workflows, coordinating up to 1,500 tool calls. The claimed benefit is a reduction in execution time by up to 4.5x compared to a single-agent approach. One demonstrated task involved autonomously using 100 sub-agents to research and identify top YouTube creators across 100 different niches.

Third is a focus on office productivity. The model is designed to handle end-to-end knowledge work, reasoning over large documents and using tools to output formatted spreadsheets, presentations, and PDFs. Internal benchmarks cited show improvements over the previous model on expert productivity tasks.

Kimi K2.5 is available through several methods: the Kimi.com website, the Kimi app, an API, and a dedicated coding product called Kimi Code. The web and app interfaces offer different modes, including an instant response option, a "thinking" mode, a standard agent mode, and the beta Agent Swarm mode.

The release positions K2.5 as a model designed for real-world, complex tasks under practical constraints.