MedGemma 1.5 4B: Multimodal Open Medical VLM

Model Overview

MedGemma 1.5 4B (google/medgemma-1.5-4b-it) is an open-weight, multimodal vision-language model developed by Google DeepMind and Google Research as part of the Health AI Developer Foundations (HAI-DEF) program.

Built upon the Gemma 3 architecture and integrating MedSigLIP (a specialized medical vision encoder), MedGemma 1.5 4B natively processes 2D/3D radiology scans (CT/MRI) and multi-patch whole-slide histopathology images (WSI), running efficiently on workstation or local edge environments.

---

Key Features

• High-Dimensional Medical Imaging Support: Native processing of 3D volumetric CT/MRI scans and whole-slide histopathology alongside 2D clinical imaging.
• Longitudinal Visual Reasoning: Compares sequential medical images (prior vs current X-rays) to track disease progression.
• Clinical Record Structuring: Extracts structured FHIR and SOAP data from unstructured physician notes and lab reports.
• MedSigLIP Integration: Medically tuned SigLIP vision encoder optimized across radiology, pathology, and dermatology.
• Compute-Efficient Deployment: 4B parameter footprint optimized for low-latency inference on local workstations and Apple Silicon (MLX).

---

Verified Project Links

• arXiv Paper: https://arxiv.org/abs/2604.05081
• GitHub Repository: https://github.com/Google-Health/medgemma
• Hugging Face Model: https://huggingface.co/google/medgemma-1.5-4b-it

---

Benchmarks & Performance

• MedQA (USMLE 4-option): 64.4% (up from 50.7% in MedGemma 1 4B).
• MedMCQA: 55.7%.
• PubMedQA: 73.4%.