Google's MedGemma: open 4B, 27B medical VLMs in Nature Medicine

Google has published its medical vision-language model family, MedGemma, in Nature Medicine, releasing weights for a 4-billion and 27-billion-parameter multimodal model, a 27B text-only variant, and MedSigLIP, a 400-million-parameter vision encoder derived from SigLIP.

On out-of-distribution tasks the paper reports 2.6–10% improvements in medical image question answering, 15.5–18.1% in chest X-ray classification, and 10.8% in agentic evaluations compared with base Gemma 3. The vision encoder was fine-tuned on over 33 million medical image-text pairs spanning radiology, histopathology, dermatology, and ophthalmology.

"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. They also report that further fine-tuning reduces errors in electronic health record information retrieval by 50%, with pneumothorax classification and histopathology patch classification reaching parity with specialized state-of-the-art methods.

The authors also concede that "specialized models may still provide better performance or be better suited to some tasks." Three-dimensional imaging such as CT and MRI series, and modalities like genomics, remain outside the current model's scope. Two of the researchers we follow shared the paper the day it posted.