What it is
Vision-centric medical MLLM that fuses specialist 2D encoders with a native 3D encoder, and adds clinically grounded evaluation for instruction following and report generation. Instead of a single 2D-centric vision tower, it composes specialist 2D encoders with a native 3D encoder via cascaded, 2D-anchored fusion, and evaluates with instruction-following plus RoI-grounded report scoring.
Key results
- Beats open-source medical MLLMs on 20 of 24 benchmarks
- Beats GPT-5.2 and Gemini-3-Flash on 13 of 16 multimodal benchmarks
- MedIF-Bench Overall-IF: 98.9, above GPT-5.2 96.0 and Gemini-3-Flash 96.6
Medical MLLMs benefit from native 3D-aware, multi-encoder vision design and clinically grounded evaluation rather than a single generic vision tower.