Google's MedGemma lands open medical VLMs in Nature Medicine

Google's MedGemma, a family of open-weight medical vision-language models, landed in Nature Medicine with checkpoints the authors say beat similarly sized generative models across radiology, histopathology, dermatology and ophthalmology tasks.

Lead authors Andrew Sellergren and Sahar Kazemzadeh, both at Google, release four artifacts: multimodal MedGemma at 4B and 27B parameters, a text-only 27B, and MedSigLIP, a 400-million-parameter medical vision encoder derived from SigLIP.

The paper's headline deltas against the base Gemma 3 models: 2.6–10% improvement in medical image question answering, 15.5–18.1% in chest X-ray finding classification, and 10.8% on the AgentClinic benchmark. On the MedQA text exam, MedGemma scores 86.2%. The abstract concludes 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."

On chest X-ray report generation, the authors report that 81% of generated reports were "likely to result in the same or superior clinical decisions" compared with the originals. That figure comes from expert review of the text itself, not from measured patient outcomes.

Two researchers we follow had the paper in circulation on day one.