Google releases MedGemma open medical VLMs in Nature Medicine
TL;DR
- Google published MedGemma, an open family of medical vision-language models built on Gemma 3, in Nature Medicine.
- MedGemma 27B scored 86.2% on MedQA, versus 78.2% for OpenBioLLM 70B at a fraction of the parameter count.
- The release spans four models: MedGemma 4B and 27B multimodal, a 27B text-only variant, and MedSigLIP, a 400M medical vision encoder.
Google has published MedGemma in Nature Medicine, releasing an open family of medical vision-language models built on Gemma 3. The 27-billion-parameter multimodal variant scored 86.2% on MedQA, the medical question-answering benchmark, against 78.2% for OpenBioLLM 70B.
The release spans four models: MedGemma 4B and 27B multimodal, a 27B text-only variant, and MedSigLIP, a 400-million-parameter medical vision encoder. The paper reports gains over base Gemma 3 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 generated chest X-ray reports, 81% were rated as producing equivalent or superior clinical decisions. Reinforcement-learning fine-tuning on MIMIC-CXR pushed the RadGraph F1 to 30.3, which the authors describe as state-of-the-art. Two researchers in our Who's Who tracker surfaced the paper shortly after it posted.
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