Google releases MedGemma, an open medical vision-language model line

MedGemma, an open-weights family of medical vision-language models from Google and Google DeepMind built on Gemma 3, posts out-of-distribution gains of 2.6 to 10 percent on medical image question answering and 15.5 to 18.1 percent on chest X-ray finding classification against its base models, the team reports in Nature Medicine.

The lineup includes two multimodal variants at 4 billion and 27 billion parameters, a 27-billion text-only variant, and a 400-million-parameter medical image encoder called MedSigLIP. Medical data made up roughly 2 percent of the vision encoder's training mix and 12 percent of multimodal decoder pretraining, a deliberately light touch the authors say is meant to balance specialization with the base model's general capabilities.

The headline clinical result is on chest X-ray report generation: 81 percent of MedGemma's generated reports were rated as resulting in the "same or superior clinical decisions" compared with originals, alongside a RadGraph F1 of 30.3 after reinforcement-learning fine-tuning. On the MedQA text benchmark, the model reaches 86.2 percent accuracy. The paper claims MedGemma "demonstrates advanced medical understanding and reasoning across images and text" while "retaining the strong general-purpose capabilities of the Gemma base models."

The authors, led by Andrew Sellergren and Sahar Kazemzadeh at Google, frame the release as a way to "accelerate medical research and the development of downstream applications." Two researchers in our Who's Who directory shared the paper's source link on release day.