Google releases MedGemma open medical vision-language model

Google has released MedGemma, a family of open-weights vision-language models fine-tuned for medical images and text, in a Nature Medicine paper published on 6 October. The 27B variant scored 86.2 on MedQA, which the authors call "the best performance for an open model at the time of its release."

Across out-of-distribution tasks, the paper reports gains over base Gemma 3 of 2.6–10% in medical image question answering, 15.5–18.1% in chest X-ray finding classification, and 10.8% in agentic evaluations. In an expert review of generated chest X-ray reports, "81% of generated reports rated as likely to result in the same or superior clinical decisions" compared with the originals, according to the authors' review by US board-certified clinicians.

The release also includes MedSigLIP, a 400-million-parameter medical image encoder derived from SigLIP. On larger comparisons, the authors write that "MedGemma outperformed the standard Gemma 3 model and overall performed similarly to the much larger API-based models." Weights are published through Google's Health AI Developer Foundations site; the authors frame the open release as enabling "data and model sovereignty" and the "ability to run locally or offline" — framing that reads directly at hospital IT departments that cannot send protected health information to a hosted API.

Two researchers we track posted the paper to their networks within hours of its release.