Kimi K2.5 is an open-source, native multimodal agentic model from Moonshot AI that extends the Kimi K2 lineage through continual pretraining on roughly fifteen trillion mixed visual and text tokens built atop Kimi-K2-Base. The model is designed to weave vision and language understanding together with advanced agentic capability, supporting both instant and thinking modes and bridging conversational and agentic interaction styles. By training on a large and balanced mix of visual and textual data, K2.5 aims to handle tasks that require jointly interpreting images, video, and natural language while staying useful for everyday assistant-style exchanges.

For complex work, Kimi K2.5 can self-direct an agent swarm of up to one hundred sub-agents that coordinate parallel workflows across as many as 1,500 tool calls, with the swarm itself created and orchestrated by the model rather than predefined. This parallel-agent approach has been shown to cut execution time by up to 4.5x compared with a single-agent setup, and the model reports strong performance on three agentic benchmarks—HLE, BrowseComp, and SWE-Verified—at a fraction of the cost of comparable systems. K2.5 is positioned as the strongest open-source model for coding with vision, particularly front-end development, where it can turn simple prompts into complete interfaces with interactive layouts and animations and reason over images or video to improve image and video-to-code generation.