Google releases MedGemma, open medical model, in Nature

Clinicians rated 81% of chest X-ray reports written by Google's new MedGemma model as "likely to result in the same or superior clinical decisions" compared to the originals, according to a paper published in Nature Medicine on 6 October 2026.

MedGemma is a family of open-weight medical foundation models built on Google's Gemma 3 architecture. The release spans multimodal variants at 4 billion and 27 billion parameters, a 27B text-only version, and MedSigLIP, a 400-million-parameter medical image encoder.

On the MedQA text benchmark, MedGemma 27B scored 86.2%, which the authors describe as the best open-model result at release aside from DeepSeek-R1, well ahead of OpenBioLLM 70B at 78.2%. Zero-shot image work improved by 15.5 to 18.1% on chest X-ray finding classification and 2.6 to 10% on medical image question-answering versus the baseline Gemma 3. The 81% report-quality figure for chest X-rays beats the larger Med-Gemini model, which the paper puts at 73%.

The authors frame the collection as "exceeding the performance of similarly sized generative models" across text and multiple imaging domains, and argue the medical pretraining "can be more effective than fine-tuning the base Gemma 3 model for medical tasks, particularly in the setting of limited training data." They also flag a 500-fold computational cost gap between MedGemma 4B and the most expensive comparator model.

The abstract publishes no per-detector accuracy numbers, and the available summary does not describe the blinding protocol behind the 81% clinician rating.