Kimi K2.5 is a native multimodal agentic model designed to advance general agentic intelligence by integrating vision and language understanding. Built upon the Kimi K2-Base, it utilizes a trillion-parameter mixture-of-experts transformer architecture. The model is engineered to jointly optimize text and vision, ensuring these modalities enhance one another rather than competing for capacity. Its primary innovation is the Agent Swarm framework, which allows the model to dynamically decompose complex tasks into smaller, heterogeneous sub-problems and execute them concurrently, significantly improving efficiency over traditional single-agent approaches.

The model underwent continual pre-training on approximately 15 trillion mixed visual and text tokens to refine its reasoning and coding performance. By coordinating up to 100 specialized AI agents simultaneously, the system achieves a 4.5x reduction in execution latency compared to standard baselines. This parallel orchestration makes it well-suited for complex workflows requiring high-level mathematical reasoning, computer use, and multimodal analysis. As an open-weight model, it is intended to support ongoing research and the development of practical, real-world agentic systems.