MedGemma: Our most capable open models for health AI development

Source: DeepMind Date: 2025-10-25 URL: https://deepmind.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/

Summary

Google released MedGemma, an open-source health AI model collection: a 4B multimodal model (64.4% MedQA, 81% radiologist-judged chest X-ray report quality) and a 27B text+multimodal model (87.7% MedQA, within 3 points of DeepSeek R1 at one-tenth the inference cost). MedSigLIP, a 400M image encoder for classification and retrieval, is included. Training data spans chest X-rays, histopathology, dermatology, and fundus images — all described as de-identified. The 4B and MedSigLIP models run on mobile hardware.

Implications

87.7% MedQA at one-tenth DeepSeek R1’s inference cost is the health AI procurement argument. MedQA tests US Medical Licensing Exam-level knowledge. 87.7% from an open-weights model at 1/10th inference cost means health AI builders no longer need frontier-model pricing for medical knowledge tasks. That’s the argument that moves hospital procurement conversations.

81% radiologist-approval rate for chest X-ray reports is a clinical validity claim that will be tested. This is a harder claim than a benchmark score — it required actual radiologist review. If it holds in external validation, MedGemma 4B is viable for radiology report assistance in resource-constrained settings (rural hospitals, low-income health systems) where the alternative is no AI assistance at all.

Open-weights health models create a regulatory question that closed-model vendors don’t face. When MedGemma is fine-tuned by a hospital on local patient data and deployed for clinical decisions, FDA regulatory status is ambiguous in ways that a Google-hosted API is not. The same openness that makes it accessible for research creates liability uncertainty for production deployment.

Watch:

• External MedQA and clinical validation studies from independent health AI research groups — the 87.7% claim needs non-Google replication
• FDA guidance on fine-tuned open health models: the regulatory gray zone is the actual deployment blocker
• MedSigLIP adoption by medical image analysis pipelines — a strong image encoder as a standalone component may have faster adoption than the full LLM