OpenMedQ beats 562B Med-PaLM M with 3.35M open medical samples
A tiny 80x smaller model just beat Google's giant Med-PaLM M on medical VQA.
OpenMedQ is a medical vision-language model pretrained on the broadest fully-open medical mix to date—14 datasets totaling ~3.35M samples spanning pathology, radiology, microscopy, and text-only clinical QA. It achieves state-of-the-art BLEU-1 of 75.9 on PathVQA, beating Med-PaLM M variants up to 562B parameters (~80x larger), and matches the best reported VQA-MED BLEU-1 (64.5). Its vision encoder, transferred to 8 unseen medical classification benchmarks under an identical downstream recipe, achieves the highest average macro-F1 (0.757) among comparable models (BiomedCLIP 0.745, PMC-CLIP 0.745, PubMedCLIP 0.746, and a from-scratch baseline 0.616). Code and an interactive demo are publicly available.
- OpenMedQ achieves 75.9 BLEU-1 on PathVQA, surpassing 562B-parameter Med-PaLM M models (80x larger), using only 3.35M pretraining samples.
- Its vision encoder tops 8 unseen medical classification benchmarks with average macro-F1 of 0.757, beating BiomedCLIP, PMC-CLIP, and PubMedCLIP.
- All code and an interactive demo are publicly released as a reproducible baseline for medical AI research.
Why It Matters
Proves smaller, open models can outperform massive closed systems in specialized medical tasks.