medgemma-27b-it

MedGemma 27B Instruct (medgemma-27b-it) is Google's largest open medical vision-language model, built on the Gemma 3 27B decoder-only transformer with a SigLIP vision encoder pre-trained on de-identified medical imagery spanning chest X-rays, dermatology, ophthalmology, and histopathology. The instruction-tuned multimodal variant accepts interleaved text and images (896x896 resolution, 256 tokens per image) plus FHIR-based electronic health record data, and generates text outputs up to 8,192 tokens within a 128K context window.

Compared with its text-only sibling, the multimodal 27B model trades a small amount of pure-text benchmark performance for unified medical text, image, and EHR reasoning. On clinical evaluations it reaches 85.3% zero-shot MedQA, 86.2% MMLU-Med, 90.5% EHRQA, and strong multimodal results such as 90.0 macro-F1 on MIMIC-CXR classification and 70.0 tokenized F1 on SLAKE radiology VQA. Test-time scaling is applied for all 27B reported scores.

Released under the Health AI Developer Foundations license in July 2025, MedGemma 27B IT is designed as a developer starting point for clinical decision support, medical report generation, EHR question answering, and radiology VQA—not as a standalone diagnostic tool. Outputs require independent verification by qualified healthcare professionals.