Google's MedGemma open medical VLMs land in Nature Medicine
TL;DR
- MedGemma's 27B multimodal variant scores 86.2% on MedQA, which the authors call the best open-model result at release.
- Over the Gemma 3 base, MedGemma gains 2.6–10% on medical image QA, 15.5–18.1% on chest X-ray classification and 10.8% on agentic tasks.
- 81% of MedGemma's chest X-ray reports were judged likely to produce the same or superior clinical decisions as the originals.
Google's medical AI team put a family of open vision-language models for medicine through peer review in Nature Medicine. The 27B multimodal variant scores 86.2% on MedQA, which the authors describe as the best open-model result at release.
Over the Gemma 3 base, the paper reports gains 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" on out-of-distribution tasks. Two researchers we track posted the paper this week.
On chest X-ray report generation, 81% of MedGemma's reports were judged "likely to result in the same or superior clinical decisions" as the originals. 68% of generated reports from normal studies and 49% from abnormal studies were rated equal to or superior to the originals.
The release spans a 4B and 27B multimodal pair, a 27B text-only sibling, and MedSigLIP, a 400M-parameter vision encoder derived from SigLIP that the paper says achieves performance "comparable to or better than that of many specialized medical image encoders." All weights are posted openly.
The evaluations are standard benchmark scores and expert rater judgments. The paper does not publish prospective clinical trial data.
Shared on Bluesky by 2 AI experts
Originally reported by nature.com
Read the original article →Original headline: An open vision-language model for diverse medical applications - Nature Medicine