Google releases MedGemma, open medical VLMs on Gemma 3

In a 306-case evaluation drawn from MIMIC-CXR, a US board-certified cardiothoracic radiologist rated 81% of the chest X-ray reports written by Google's MedGemma 27B as clinically equivalent or superior to the human-written originals. That is one of the headline numbers in the paper published in Nature Medicine on October 6.

MedGemma is a four-model release: a 4B and 27B multimodal pair, a 27B text-only sibling, and MedSigLIP, a 400-million-parameter medical vision encoder. All four are built on Gemma 3 and shipped openly. On MedQA the 27B model hits 86.2%, which the authors flag as the "best performance for an open model at release" once 671B-parameter systems are set aside. On chest X-ray classification the paper reports gains of 15.5 to 18.1 points over the base Gemma 3.

The pitch underneath the benchmarks is cost. The paper describes "a 500-fold difference in computational cost" between MedGemma 4B and the most expensive comparator it is set against. For hospital IT groups that cannot or will not pipe patient data into a closed API, that gap is the actual argument.

The authors are explicit about where the model earns its keep: "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." After reinforcement-learning fine-tuning, chest X-ray RadGraph F1 reaches 30.3.

Caveats sit inside the numbers. The 81% equivalence rating comes from a single radiologist, not a panel. The 500-fold cost comparator is not named in the summary. And medical content was kept to 2% of the vision encoder's training mixture and 12% of multimodal decoder pretraining, which the authors frame as preserving general capability. That also means this is a specialist lightly tuned on top of a generalist, not a model built from medical data up. Two of the researchers in our tracker had already shared the paper by the time it reached us.