Google's MedGemma lands open medical VLMs in Nature Medicine

Google has released MedGemma, a family of open medical vision-language foundation models built on Gemma 3, with the training and evaluation published in Nature Medicine on October 6. The release contains two multimodal variants at 4B and 27B parameters, plus MedSigLIP, a medically tuned vision encoder derived from SigLIP.

The authors frame the problem flatly in the abstract: "Artificial intelligence has high potential for impact in healthcare applications, but its training and deployment are challenging due to diverse data, a complex spectrum of possible tasks and important privacy needs." Their claim for the family is that it is "approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models."

Against base Gemma 3 on out-of-distribution tasks, the paper reports improvements of 2.6–10% on medical image question answering, 15.5–18.1% on chest X-ray finding classification, and 10.8% on agentic evaluations. Fine-tuning on narrower subdomains stacks further gains, including a 50% reduction in errors on electronic health record information retrieval.

Two researchers we track shared the paper the day it dropped. The retrieved abstract does not publish per-detector accuracies or a regulatory-pathway section; the comparison is against base Gemma, not against the closed clinical specialists hospitals currently license.