medgemma-1.5-4b-it

MedGemma 1.5 4B IT is Google's open multimodal medical foundation model, released January 13, 2026, and built on the Gemma 3 decoder-only Transformer architecture with a SigLIP vision encoder pre-trained on de-identified medical imaging data. With 4 billion parameters and a 128K-token context window, it accepts text and images (896x896 resolution, encoded to 256 tokens per image) and generates text outputs up to 8K tokens. It is the first open-source multimodal model to natively support high-dimensional medical imaging, including 3D CT and MRI volumes, whole-slide histopathology, longitudinal chest X-ray comparison, anatomical bounding-box localization, medical document understanding, and electronic health record interpretation.

Compared to MedGemma 1 4B, version 1.5 delivers substantial gains across medical imaging and clinical reasoning tasks: 3D MRI condition classification improves by 11 percentage points, 3D CT by 3 points, whole-slide pathology macro F1 by 47 points, chest X-ray anatomical localization IoU by 35%, MedQA accuracy rises from 64.4% to 69.1%, and EHRQA accuracy jumps from 67.6% to 89.6%. The model is evaluated on clinically relevant benchmarks spanning radiology, dermatology, histopathology, ophthalmology, and medical text reasoning, though it shows a modest trade-off on general benchmarks such as MMLU-Pro (33.8% vs. 39.1% for MedGemma 1 4B).

MedGemma 1.5 is intended as a developer starting point for healthcare AI applications rather than a standalone clinical diagnostic tool. All outputs require independent verification by qualified medical professionals. The model is distributed under the Health AI Developer Foundations terms of use via Hugging Face and is designed for efficient local deployment, fine-tuning, and integration with companion tools such as MedASR for medical speech-to-text workflows.