Google releases MedGemma, open medical VLMs built on Gemma 3

Google's health AI team released MedGemma, a family of open medical vision-language models that scored 86.2% on MedQA and outperformed baseline Gemma 3 by 15.5 to 18.1 percentage points on chest X-ray finding classification, according to a paper in Nature Medicine.

The collection includes two multimodal models at 4 billion and 27 billion parameters, a 27B text-only variant, and a 400M-parameter vision encoder called MedSigLIP, all built on Gemma 3. On a chest X-ray report-generation task the paper puts the RadGraph F1 at 30.3, and reports that 81% of generated reports were rated as producing equivalent or superior clinical decisions versus ground truth.

The result most likely to get argued over sits in the agentic section: on the AgentClinic benchmark, the 27B model exceeded human physician performance, the paper states. Lead authors Andrew Sellergren and Sahar Kazemzadeh, with roughly 90 collaborators credited across Google Research and Google DeepMind, pitch the release as scaffolding rather than a finished product, writing that the collection "provides a strong foundation of medical image and text capabilities, with the potential to accelerate medical research."

Benchmark SOTA is not a license to deploy in a ward. The 81% equivalence figure means roughly one in five generated reports was judged inferior to ground truth, and AgentClinic is a benchmark, not a clinical setting.