Google publishes MedGemma, open medical vision-language model

A Google-led team has published MedGemma, a collection of open medical vision-language foundation models built on Gemma 3, in Nature Medicine. The authors report out-of-distribution gains of 2.6-10% on medical image question answering, 15.5-18.1% on chest X-ray finding classification, and 10.8% on agentic evaluations against the base Gemma 3 models.

The paper also introduces MedSigLIP, "a medically tuned vision encoder derived from SigLIP" that the authors say reaches performance comparable to or better than many specialized medical image encoders. MedSigLIP powers the visual side of MedGemma.

The authors frame the problem the collection is meant to address as a structural one. "Artificial intelligence has high potential for impact in healthcare applications, but its training and deployment are challenging due to diverse data, a complex spectrum of possible tasks and important privacy needs," the abstract reads. Their practical claim is that fine-tuning MedGemma can be more effective than fine-tuning the base Gemma 3 model, "particularly in the setting of limited training data", the regime most hospital research groups actually sit in.

The abstract reports no head-to-head numbers against established specialized medical models and no clinical validation data; the comparisons are to the general-purpose Gemma base.