Google releases MedGemma, open medical vision-language models

Google has released MedGemma, a family of open vision-language models tuned for medicine, in a paper in Nature Medicine. The release includes two multimodal weights, 4 billion and 27 billion parameters, a 27B text-only variant, and a 400-million-parameter medical vision encoder called MedSigLIP, all built on Gemma 3.

On the headline text benchmark the paper reports MedQA accuracy of 86.2%, described as the best open model performance at release. For out-of-distribution tasks, the paper says "improvements ranged from 2.6–10% in medical image question answering to 15.5–18.1% in chest X-ray findings classification," measured against baseline Gemma 3.

The chest X-ray claim: "68% of generated reports from normal studies and 49% from abnormal studies rated equal to or superior to original reports." The gap between normal and abnormal cases sits in the numbers; the abstract does not name the adjudicators or the blinding protocol.

MedSigLIP is positioned as a drop-in medical vision backbone. The paper says the 400-million-parameter encoder "achieves performance comparable to specialized medical image encoders" across radiology, histopathology, dermatology and ophthalmology.

Co-first authors Andrew Sellergren and Sahar Kazemzadeh led the work, with contributors drawn across Google Research and Google DeepMind.