An open vision-language model for diverse medical applications

MedGemma is an open collection of medically tuned vision-language foundation models built on the Gemma 3 architecture, offered in multimodal 4-billion and 27-billion parameter variants plus a 27-billion text-optimized variant. Training used millions of medical image-text pairs to create MedSigLIP, a 400-million-parameter medical image encoder that powers visual understanding. The models are designed to retain Gemma 3’s general-language strengths while adding domain-specific capabilities for radiology, pathology, dermatology and electronic health record question answering. MedGemma emphasizes size-to-performance efficiency, enabling data-efficient fine-tuning for diverse downstream clinical tasks.

Across extensive benchmarks, MedGemma outperforms or matches much larger proprietary models and specialized systems: out-of-distribution gains reported include 2.6-10% improvement in medical visual question answering, 15.5-18.1% in chest X‑ray finding classification and a 10.8% gain in agentic evaluations versus base models. Text benchmarks (MedQA, MedMCQA, PubMedQA, MMLU Med, AfriMed‑QA, AgentClinic) show superior results to untuned Gemma 3 and competitive results with frontier models. MedSigLIP alone delivers comparable or better zero-shot and few-shot image classification and retrieval than many specialist encoders. The open release aims to accelerate medical AI research by enabling researchers to fine-tune and adapt these models for specific clinical workflows.