Google releases MedGemma, open medical vision-language models

Google has published MedGemma, a collection of openly released medical vision-language models built on Gemma 3, in Nature Medicine.

The abstract describes the collection as models that "demonstrate 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." It is paired with MedSigLIP, "a medically tuned vision encoder derived from SigLIP," that powers the visual side.

On out-of-distribution tasks, the authors report gains 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 compared with the base models."

MedSigLIP, the paper says, "achieves performance comparable to or better than that of many specialized medical image encoders."

The authors flag their own limits. "Specialized models may still provide better performance or be better suited to some tasks," they write, and warn that "the potential for model errors or subtle forms of distribution bias in training data remains an important aspect to address." Thoughtful validation of safety, performance and reliability, they add, "remains a critical aspect" for downstream applications.

Two researchers in our Who's Who tracker circulated the paper on release.