Open foundation models for clinical AI

What happened

On November 12, 2026, researchers led by Andrew Sellergren and Lin Yang released MedGemma, a set of open medical foundation models. MedGemma includes both multimodal and text-only models, trained on a wide range of clinical data types such as radiology, pathology, dermatology, and ophthalmology. The models were evaluated on several benchmarks and clinical reasoning tasks, where they surpassed previous models. The release aims to support various healthcare AI applications by enabling adaptation to different clinical environments.

Key details

• MedGemma includes multimodal models with 4 billion and 27 billion parameters, and a text-only model with 27 billion parameters.
• A separate medical image encoder called MedSigLIP is part of the release.
• The models were trained using data from multiple medical specialties, including radiology and pathology.
• In the AgentClinic simulated environment, MedGemma 27B performed well on longitudinal clinical reasoning tasks.
• Fine-tuning MedGemma 4B with just 10% of task-specific data outperformed the similarly fine-tuned Gemma 3 model.

Why it matters

Healthcare AI developers and clinical researchers now have access to open models that can be adapted to local healthcare settings and integrated with clinical systems. This enables more efficient use of limited medical data and supports complex reasoning across diverse patient records. Hospitals and research institutions can use these models to improve the performance of AI tools in real-world clinical environments while reducing the amount of data needed for customization.