Open source AI model accelerates progress in medical video analysis

United Imaging Intelligence (UII) has initiated a project to foster advancements in artificial intelligence (AI) for medical video analysis. The company has released an open-source model, a public dataset, and a benchmark to provide researchers with a common foundation for collaboration.

Understanding medical videos presents significant challenges in AI, requiring precise spatial perception and complex temporal reasoning. Progress has often been hindered by a lack of clinical data and the high cost of expert annotation. UII aims to overcome these obstacles with an open framework designed to measure research progress and facilitate international cooperation.

Following the acceptance of its MedGRPO research at CVPR 2026, UII launched its uAI NEXUS MedVLM model and made the MedVidBench test dataset, comprising over 6,200 samples, publicly available. These are part of a larger collection containing over 530,000 video-instruction pairs. Additionally, UII introduced the public MedVidBench leaderboard to evaluate models across ten metrics.

To translate this shared foundation into collective progress, UII co-launched the MedVidU Challenge in Summer 2026 with Université de Strasbourg and Technical University of Munich. The competition attracted 75 teams from 18 countries and regions, aiming to gather diverse ideas and approaches for AI in medical video.

The MedVidBench dataset has exceeded 30,000 downloads and has garnered interest from the international research community. Four teams advanced to the final stage of the MedVidU Challenge, with two teams presenting their findings at the ECCV 2026 MedVidU Workshop in Malmö, Sweden, in September. This initiative promotes an open ecosystem combining resource sharing, transparent evaluation, and research exchange to advance AI in medical video towards clinical practice.