MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images

achieves an average improvement of +6.6% over Fleming-VL-8B, with gains of +6.0% on MMMU-Med, +9.8% on PMC-VQA, and +21.3% on MedXpertQA. For text-based QA, it attains +14.4% over Fleming-VL-8B, driven by +8.4% on MMLU-Med and +30.1% on MedQA. In medical report generation, MedMO-8B-Next delivers +6.7% on MIMIC-CXR. Moreover, it exhibits strong grounding capability with a Bacteria IoU of 56.1, representing a +47.8 IoU gain over Fleming-VL-8B, underscoring its robust spatial reasoning and localization performance. MedMO-4B-Next remains highly competitive at its smaller scale, surpassing Fleming-VL-8B across VQA, QA, and report generation benchmarks. Evaluations across radiology, ophthalmology, and pathology microscopy confirm MedMO's broad cross-modality generalization.

We further conduct comprehensive experiments and analyses on data curation, training, and alignment strategies, providing a transparent and reproducible framework for future medical MLLM development. Extensive evaluations demonstrate that MedMO achieves state-of-the-art (SOTA) performance across diverse benchmarks, surpassing prior open and proprietary systems on tasks including medical VQA, report generation, and diagnostic reasoning.

Our main contributions are summarized as follows:

• We develop a powerful open-source post-trained multimodal large VLM, MedMO, designed for comprehensive medical image understanding and grounding.

• We curate over 26M multimodal medical and biomedical samples from 45 datasets and establish a multi-stage posttraining that progressively enhances cross-modal alignment and reasoning. This provides a scalable roadmap toward a generalist foundation model for medical.

• To evaluate VLM performance on detection tasks, we construct a dedicated Cell dataset from opensource microscopy images with varying sizes, shapes, and densities.

• We conduct extensive experiments and analyses across data and methodology dimensions, providing an open benchmark for future multimodal medical LLM research and training recipes.