Google publishes MedGemma open medical VLMs in Nature Medicine

MedGemma, a collection of open medical vision-language models from Google built on Gemma 3, scored 86.2% on MedQA, which the paper calls "the best performance for an open model" at release.

The work landed in Nature Medicine on October 6, 2026, alongside MedGemma 4B and 27B multimodal variants, a 27B text-only variant, and MedSigLIP, a 400-million-parameter medical vision encoder. On out-of-distribution tasks, 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" over the base Gemma models. Fine-tuning on MIMIC-CXR produced a RadGraph F1 of 30.3 with reinforcement learning.

The clinical read-out is narrower than the benchmark sweep. A US board-certified cardiothoracic radiologist reviewed 306 MedGemma-generated chest X-ray reports drawn from MIMIC-CXR and judged 81% to produce "same or superior clinical decisions" versus the originals.

"High-performing foundation models that enable data-efficient fine-tuning for diverse downstream tasks can meaningfully accelerate development in this domain," the authors write. The evaluated domains span radiology, histopathology, dermatology and ophthalmology.

Two of the researchers we follow in our Who's Who directory posted the paper the day it went up.