Kimi-K2.5

Kimi-K2.5 is a multimodal agentic model engineered to advance general agentic intelligence by seamlessly integrating vision and language processing. Built upon a trillion-parameter mixture-of-experts transformer architecture, the model is designed to treat text and visual inputs as mutually enhancing components rather than competing data streams. Its primary innovation is the Agent Swarm framework, which allows the model to dynamically decompose complex, heterogeneous problems into smaller sub-tasks. By executing these tasks concurrently, the architecture significantly reduces latency compared to traditional single-agent approaches, making it highly effective for sophisticated coding, reasoning, and computer-use applications.

The model lineage stems from continual pre-training on approximately 15 trillion tokens, which establishes a robust foundation for its specialized operational modes. Through this training, Kimi-K2.5 achieves strong performance on benchmarks like Humanity’s Last Exam, demonstrating high efficiency in both execution speed and cost. Its design supports a variety of paradigms, including instant and thinking modes, which allow users to tailor the model to specific conversational or agentic requirements. By providing this capability as an open-weight model, it serves as a versatile tool for researchers and developers looking to deploy scalable, parallelized agentic systems in real-world production environments.