UII Open-Sources uAI NEXUS MedVLM, a 4B/7B-Parameter Medical Video LLM, Plus the MedVidBench Benchmark
- United Imaging Intelligence released uAI NEXUS MedVLM, an open-source medical video LLM with 4B and 7B parameter versions, trained on 531,850 video-instruction pairs covering 8 clinical scenarios such as robotic surgery, endoscopy, and nursing care.
- On surgical safety assessment the model reaches 89.4% accuracy, against 1.8% for GPT-5.4 and 10.1% for Gemini 3.1, and it scores 4.2 of 5 on video report generation where GPT-5.4 scores 2.5 and Gemini 3.1 scores 2.4.
- In spatio-temporal action localization it reports up to 14x higher mIoU than GPT-5.4 and 4x higher than Gemini 3.1, with numbers drawn from the paper at arXiv 2512.06581 and accepted by CVPR 2026.
- UII also released 6,245 test samples from its MedVidBench benchmark spanning 8 surgical datasets, with a leaderboard that scores submissions against private ground truth and updates a global ranking continuously.
- The model does spatio-temporal localization of instruments, procedural recognition, structured clinical report generation, next-step prediction, and surgical skill and safety risk assessment from video.