Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model (www.kimi.com)

Kimi has announced the release of Kimi K2.5, its most advanced open-source multimodal model, which enhances the capabilities of its predecessor, Kimi K2, through continued pretraining on around 15 trillion mixed visual and textual tokens. K2.5 introduces a self-directed agent swarm paradigm capable of orchestrating up to 100 sub-agents to execute complex workflows in parallel. This innovation results in potential execution time reductions of up to 4.5 times compared to traditional single-agent setups, marking a significant leap in efficiency for complex tasks. The K2.5 model excels in both coding, particularly in front-end development, and in visual reasoning, enabling it to convert simple conversations into interactive web interfaces or reconstruct websites from video content. Built on a foundation of Parallel-Agent Reinforcement Learning (PARL), K2.5 dynamically instantiates agents to tackle distributed tasks more efficiently. With new benchmarks showcasing substantial improvements over previous iterations and advanced features like Kimi Code for integrating with IDEs, Kimi K2.5 positions itself as a pivotal tool in the AI/ML community, driving forward the frontier of agentic intelligence in real-world applications and productivity.