MedGemma, Google's open medical VLM, scores 86.2 on MedQA
TL;DR
- MedGemma scored 86.2 on MedQA, which the paper calls the best performance for an open model.
- Out-of-distribution gains over base Gemma 3: 2.6–10% on medical image QA, 15.5–18.1% on chest X-ray findings, 10.8% on agentic evaluations.
- MedSigLIP, a medically tuned vision encoder derived from SigLIP, ships alongside and matches or beats specialized medical image encoders.
Google's MedGemma scored 86.2 on MedQA, which the paper calls "the best performance for an open model." The release is a collection of medical vision-language foundation models built on Gemma 3.
On out-of-distribution tasks, the authors report gains over the base model 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." Alongside the main models, the paper introduces MedSigLIP, "a medically tuned vision encoder derived from SigLIP," which the authors say "achieves performance comparable to or better than that of many specialized medical image encoders."
The pitch is efficient fine-tuning. "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 authors write.
The paper is cautious about clinical deployment. "Thoughtful validation of safety, performance and reliability for downstream applications remains a critical aspect," they add.
Two of the researchers we follow flagged the paper the day it landed.
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