Google openly releases MedGemma medical vision-language models

Google's medical AI team has published a paper in Nature Medicine describing MedGemma, a family of medical vision-language models built on Gemma 3 and released openly. On MedQA, the paper reports 86.2% accuracy. On chest X-ray classification, it claims a 15.5–18.1% improvement over base Gemma 3; on medical image question answering, 2.6–10%.

The headline clinical number comes from radiology report generation. "81% of generated reports were rated as likely to result in the same or superior clinical decisions," the paper reports, based on expert clinician review. The RadGraph F1 score reached 30.3 with reinforcement-learning fine-tuning.

The release spans four artifacts: MedGemma 4B and 27B multimodal, a 27B text-only version, and MedSigLIP, a 400M-parameter medically tuned vision encoder. The authors argue the small encoder "performs comparably to specialized medical encoders," which matters for hospitals that want to run inference locally rather than route images through an API.

"MedGemma demonstrates advanced medical understanding and reasoning across images and text and multiple medical imaging domains, exceeding the performance of similarly sized generative models while maintaining the general capabilities of the Gemma base models," the authors write. On general benchmarks like MMLU Pro and MMMU, they report "only minor decreases" versus base Gemma 3.

Training mixed real medical imaging (chest X-rays, CT/MRI 2D slices, roughly 32.5 million histopathology patches, dermatology images covering 210+ skin conditions, retinal images) with text sets including MedQA, MedMCQA and PubMedQA plus around 200,000 synthetic medical questions. Medical data was about 12% of the multimodal decoder training mix.

Fine-tuning efficiency is the operational pitch. With 10% of task data available, the paper reports MedGemma 4B "substantially outperformed Gemma 3 4B" on pneumothorax, histopathology and skin-lesion classification.

Comparison claims are hedged. "MedGemma outperformed the standard Gemma 3 model and overall performed similarly to much larger API-based models," the authors write. The paper also reports MedGemma exceeded human physician performance on the AgentClinic-MedQA environment, a simulated clinical benchmark rather than a prospective patient-care study.

The authors say they "openly released MedGemma and MedSigLIP to facilitate their widespread evaluation, improvement and adaptation by the community." A pair of researchers we follow passed the Nature link around within hours of publication.