United Imaging Intelligence Releases Open Source Medical Video Large Language Model

United Imaging Intelligence has released uAI NEXUS MedVLM, described as the first open source medical video large language model, announced in a press release. The model is designed for clinical video understanding and has been accepted by CVPR 2026, a major computer vision conference.

uAI NEXUS MedVLM is trained on 531,850 video-instruction pairs from eight medical scenarios, including robotic and open surgery, endoscopy, and nursing care. With 4 billion and 7 billion parameter versions, it surpasses general foundation models such as GPT-5.4 and Gemini 3.1 in key medical video tasks. The model achieves 89.4 percent accuracy in surgical safety assessment and significantly higher performance in spatio-temporal action localization and video report generation.

To support further research, United Imaging Intelligence has introduced MedVidBench, a benchmark dataset containing over six thousand annotated test samples across eight surgical datasets. The dataset enables automatic evaluation through a public leaderboard with private ground truth comparisons. Researchers and healthcare institutions are invited to participate in this open challenge to advance medical video intelligence.

The company stated that uAI NEXUS MedVLM integrates perception, reasoning, and decision-making for precise spatio-temporal localization and procedural recognition. It can generate structured clinical reports and assess surgical skills, forming a foundation for future AI systems in healthcare automation.