Google open-sources MedGemma medical vision-language models

Google researchers have released MedGemma, an open family of medical vision-language models built on the Gemma 3 architecture, with the full write-up in Nature Medicine. The release bundles a 4-billion-parameter multimodal variant, a 27-billion-parameter multimodal variant, a 27-billion-parameter text-only variant, and a 400-million-parameter medical vision encoder called MedSigLIP.

On out-of-distribution tasks, the paper reports MedGemma gains "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 the base models." On MedQA, the model scored 86.2%, the best open-model result at the time of release. In an expert read of chest X-ray reports generated by the model, 81% produced clinical decisions rated equivalent to or better than those from the original radiologist-written reports.

The authors argue a generalist model is not enough for the clinic. "While general-purpose (nonmedically tuned) large multimodal models demonstrate impressively broad abilities, generic models can lack the nuanced medical understanding needed to interpret and reason about medical data in a robust way," they write. The team, led by first author Andrew Sellergren with joint supervision from Shravya Shetty, David F. Steiner, Rory Pilgrim, Daniel Golden and Lin Yang, all affiliated with Google, say they "openly released MedGemma and MedSigLIP to facilitate their widespread evaluation, improvement and adaptation by the community across multiple modalities."

Caveats run through the paper. The authors flag possible train-test contamination on public benchmarks, histopathology performance that still trails specialized encoders, and no 3D handling for full CT or MRI series. They note the models still need validation before real-world healthcare deployment.

Fine-tuning the 4B variant with reinforcement learning on MIMIC-CXR hit a state-of-the-art RadGraph F1 of 30.3, which the authors report alongside a roughly 500-fold computational cost advantage over the most expensive comparator model they tested.