Google's MedGemma debuts as open medical vision-language model

Google has published MedGemma, an open-weights family of medical vision-language models built on Gemma 3, in Nature Medicine. On out-of-distribution medical tasks, the paper reports gains of 2.6 to 10 percent in medical image question answering, 15.5 to 18.1 percent in chest X-ray finding classification, and 10.8 percent in agentic evaluations compared with the base models. MedGemma reaches 86.2 percent accuracy on MedQA. For chest X-ray report generation on MIMIC-CXR, the authors report a RadGraph F1 of 30.3 with reinforcement-learning fine-tuning.

The release has three generative pieces: a 4-billion-parameter multimodal variant, a 27-billion-parameter multimodal variant, and a 27-billion-parameter text-only variant. Alongside them sits MedSigLIP, a 400-million-parameter medical vision encoder derived from SigLIP, which the paper says achieves performance "comparable to or better than that of many specialized medical image encoders."

The authors frame the release as community infrastructure. "We openly released MedGemma and MedSigLIP to facilitate their widespread evaluation, improvement and adaptation by the community," they write. Fine-tuning MedGemma, the paper argues, can be more effective than fine-tuning the base Gemma 3 model for medical tasks, "particularly in the setting of limited training data."

They are also direct about the limits of their own numbers. "Limitations exist for these benchmarks such that evaluating performance on them represents only the first step towards validating real-world utility," the paper states, adding that model errors and subtle forms of distribution bias in training data remain an important aspect to address. Lead authors Andrew Sellergren and Sahar Kazemzadeh, both at Google in Mountain View, report a 500-fold difference in computational cost between MedGemma 4B and the most expensive comparator model. Two of the researchers on our radar shared the paper after it posted.