Google releases MedGemma medical vision-language models

Google published MedGemma in Nature Medicine on 6 October 2026: a family of medical vision-language models built on Gemma 3, released openly in 4B and 27B parameter configurations.

The claims are relative, not absolute. Compared with the base Gemma 3 models, the authors report "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." The paper also ships MedSigLIP, a 400M-parameter vision encoder derived from SigLIP that the authors say "achieves performance comparable to or better than that of many specialized medical image encoders."

The pitch is to downstream developers. The paper argues that 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." Two researchers we track shared the paper the week it ran.

The authors also claim a "500-fold difference in computational cost between MedGemma 4B and the most expensive comparator model." The abstract reports only percentage-point deltas; no absolute per-benchmark scores appear in it.