Google's MedGemma: open medical vision-language models in Nature

Google researchers describe MedGemma in Nature Medicine, a collection of medical vision-language foundation models built on Gemma 3. The release spans a 4-billion-parameter multimodal variant, a 27-billion-parameter multimodal variant and a 27B text-only version, along with a medically tuned vision encoder called MedSigLIP derived from SigLIP.

The headline numbers come from out-of-distribution tasks. The paper reports 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 the base models." On chest X-ray report generation, 81% of the reports MedGemma produced were rated as "likely to result in same or superior clinical decisions."

The pitch is laid out plainly. "Generic models can lack the nuanced medical understanding needed to interpret and reason about medical data in a robust way," the authors write, arguing 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."

The cost picture is the part that most changes who can play. The paper reports "a 500-fold difference in computational cost between MedGemma 4B and the most expensive comparator model."