Google's MedGemma open medical VLMs land in Nature Medicine
TL;DR
- MedGemma 27B's chest X-ray reports were rated clinically equal or superior to the originals in 81% of a 306-case MIMIC-CXR evaluation.
- Against base Gemma 3, MedGemma gains 2.6–10% on medical image question answering and 15.5–18.1% on chest X-ray finding classification.
- The release ships 4B and 27B multimodal checkpoints, a text-only 27B, and MedSigLIP, a 400M medical vision encoder derived from SigLIP.
MedGemma 27B's generated chest X-ray reports were rated clinically equivalent or superior to the originals in 81% of a 306-case evaluation drawn from MIMIC-CXR, with the grading done by a US board-certified cardiothoracic radiologist. The headline number, published in Nature Medicine, hides a sharper split underneath it: 68% on normal studies, 49% on abnormal.
Google is also comparing MedGemma against its own prior work. The paper reports that 73% of chest X-ray reports from the larger Med-Gemini would have led to the same or better clinical decisions than the original, so the smaller, newer MedGemma edges Med-Gemini on this read.
The abstract is plain on the base-model lift: "MedGemma achieves 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 MedQA, the paper reports 86.2% accuracy for MedGemma 27B.
The release is a family. MedGemma 4B and 27B multimodal checkpoints, a text-only 27B, and MedSigLIP, a 400-million-parameter medical vision encoder derived from SigLIP. Lead authors Andrew Sellergren and Sahar Kazemzadeh are at Google, with roughly 90 collaborators credited.
Two of the researchers we follow had the paper circulating in their feeds the week it went live.
Shared on Bluesky by 2 AI experts
Originally reported by nature.com
Read the original article →Original headline: An open vision-language model for diverse medical applications - Nature Medicine