MedGemma: Google's open medical vision-language model family

Google has released MedGemma, a family of open medical vision-language models built on its Gemma 3 architecture, in a research article published in Nature Medicine. The 4B and 27B parameter variants cover radiology, histopathology, dermatology and ophthalmology, with the authors reporting improvements of 2.6–10% in medical image question answering, 15.5–18.1% in chest X-ray classification, and 10.8% in agentic evaluations against the base Gemma 3 models.

On the AgentClinic benchmark, 'MedGemma 27B exceeded human physician performance on AgentClinic-MedQA,' the paper reports, and the 27B text variant hit 86.2% on MedQA, described as the best score of any open model at release. For MIMIC-CXR chest X-ray reporting, blinded review rated 68% of generated normal-study reports equal or superior to the originals, and 81% likely to result in the same or better clinical decisions. Two of the medical-AI researchers we track had circulated the paper early in our Who's Who feed.

The authors, led by Andrew Sellergren and Sahar Kazemzadeh at Google, write that 'the open nature of these models is intended to enable researchers worldwide to continue improving them for specific needs.' The release also includes MedSigLIP, a 400-million-parameter vision encoder pitched as comparable to specialized medical image encoders.

The headline numbers come from retrospective benchmarks rather than prospective clinical validation, and the 68–81% MIMIC-CXR review ratings leave roughly a fifth to a third of generated reports rated worse than the originals.