An open vision-language model for diverse medical applications.

MedGemma, a collection of medical vision-language foundation models built on Gemma 3, outperforms similarly sized generative models across medical images and text. On out-of-distribution tasks it improves medical image question answering by 2.6-10%, chest X-ray finding classification by 15.5-18.1%, and agentic evaluations by 10.8% over base models. Its medically tuned vision encoder, MedSigLIP, matches or beats many specialized medical image encoders, and fine-tuning MedGemma beats fine-tuning Gemma 3 directly, especially with limited training data.