Google open-sources MedGemma medical vision-language models

Google researchers published MedGemma in Nature Medicine on October 6, releasing a family of open medical vision-language models built on Gemma 3. The collection spans four sizes: a 4-billion-parameter multimodal model, a 27-billion-parameter multimodal model, a 27B text-only variant, and MedSigLIP, a 400-million-parameter medical vision encoder.

The paper reports 86.2% accuracy on MedQA, with gains of 15.5–18.1% in chest X-ray finding classification, 2.6–10% in medical image question answering, and 10.8% in agentic evaluations relative to the base Gemma models. For radiology report generation on MIMIC-CXR, MedGemma achieves a RadGraph F1 of 30.3 with RL fine-tuning, and the authors report that 81% of generated reports were rated likely to produce equivalent clinical decisions.

Training used extended internal corpora including 184,852 additional retina images, 51,049 dermatology images across 210 skin conditions, and roughly 32.5 million histopathology patch-text pairs.

Lead authors Andrew Sellergren and Sahar Kazemzadeh write that the model "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 release goes out openly through Google Health AI Developer Foundations.