Kimi has announced the release of its first-generation open-source model, Kimi K2.5, through its official WeChat account. According to Foresight News, Kimi K2.5 has achieved the highest scores among global open-source models in various agent evaluations, including HLE (Human Last Examination), BrowseComp, and DeepSearchQA.
The K2.5 model is designed with a native multimodal architecture, supporting both visual and text inputs. It integrates capabilities such as visual understanding and reasoning, programming, and agent functionalities into a single model. The release and open-sourcing of the Kimi K2.5 model introduce new capabilities in visual comprehension, coding, and agent clustering.