Google's open MedGemma trails only DeepSeek-R1 on MedQA

MedGemma scored 86.2% on MedQA, bested only by DeepSeek-R1 at 671 billion parameters among open models on the medical question-answering benchmark. Google published the family in Nature Medicine this month: four variants built on Gemma 3, consisting of a 4B multimodal model, a 27B multimodal model, a 27B text-only model, and MedSigLIP, a 400-million-parameter vision encoder.

Against base Gemma 3 on out-of-distribution tasks, the paper reports improvements of "2.6–10% in medical image question answering, 15.5–18.1% in chest X-ray finding classification and 10.8% in agentic evaluations." On MIMIC-CXR chest X-ray report generation with reinforcement learning fine-tuning, the authors report a state-of-the-art RadGraph F1 of 30.3, with reviewers rating "68% of normal reports and 49% of abnormal reports" as equal or superior to originals.

"500-fold difference in computational cost between MedGemma 4B and the most expensive comparator model," the authors note.

The recipe stays restrained. Medical data was kept to 2% of vision encoder training and 12% of multimodal decoder pretraining, enough to specialize without destroying general capability. The authors flag that benchmarks "may be approaching saturation," that specialized models still outperform MedGemma on some histopathology tasks, and that 3D medical imaging such as CT and MRI series is not supported.

Models are released through Google's Health AI Developer Foundations site. The work was jointly supervised by Shravya Shetty, David F. Steiner, Rory Pilgrim, Daniel Golden and Lin Yang, with over 100 Google contributors listed. Two clinician-researchers in our Who's Who directory shared the paper's link.