Google releases MedGemma, open medical vision-language model

Google and Google DeepMind have published MedGemma in Nature Medicine, 'a collection of medical vision-language foundation models based on Gemma 3' designed to interpret and reason over medical images and text.

Against the base Gemma 3 models, 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.' On MedQA, 'MedGemma's performance of 86.2 on MedQA represented the best performance for an open model at the time of its release,' the authors write.

Alongside the generative models the paper introduces MedSigLIP, which it describes as 'a medically tuned vision encoder derived from SigLIP' that 'achieves performance comparable to or better than that of many specialized medical image encoders.'

The paper's data-efficiency claim is where it lands the sharpest line: '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.' Licensing details and per-variant parameter counts are not stated in the abstract.