Google releases MedGemma, open medical VLMs built on Gemma 3

Google's Nature Medicine paper introduces MedGemma, a collection of open medical vision-language models built on Gemma 3 that the authors report beat similarly sized generative models on medical tasks while retaining general capabilities. The lineup includes a 4B and 27B multimodal model, a 27B text-only variant, and MedSigLIP, a 400-million-parameter medical vision encoder.

The headline numbers: on out-of-distribution medical tasks, MedGemma shows improvements 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 compared with base models." On MedQA it reached 86.2% accuracy, which the paper describes as the best open model performance at release.

For radiology report generation on MIMIC-CXR chest X-rays, 81% of generated reports were rated likely to lead to the same or superior clinical decisions, with 68% of normal and 49% of abnormal studies judged equal or better than human-written originals. The RL-fine-tuned variant hit a RadGraph F1 of 30.3. On AgentClinic-MedQA, the authors report that MedGemma 27B exceeded human physician performance.

The authors argue the real pull is fine-tuning efficiency, not raw leaderboard numbers. "Fine-tuning MedGemma can be more effective than fine-tuning the base Gemma 3 model for medical tasks, particularly in the setting of limited training data," the paper states. With just 10% of task-specific data, MedGemma 4B outperformed the Gemma 3 4B baseline on pneumothorax detection, histopathology (CRC100k) and skin-lesion classification (ISIC 2017).

Training leaned on roughly 32.5 million histopathology patch-text pairs, 51,049 dermatology images across 210 skin conditions, and 184,852 retinal images, alongside MIMIC-CXR, NIH ChestX-ray and other radiology and EHR sources. Weights, code and evaluation harnesses are released on Google's Health AI Developer Foundations site. Two researchers we follow shared the paper on the day it posted.

The paper is upfront about what MedGemma does not do. It does not process 3D CT or MRI volumes, it lacks genomics, and the authors concede that specialized single-task models can still outperform it on narrow tasks.