MedGemma

MedGemma is a family of open medical multimodal models released by Google in July 2025 and described in the MedGemma Technical Report. Built on the open Gemma 3 architecture, it specializes the base models for medical text and image comprehension while keeping the weights downloadable, distinguishing it from Google's earlier Med-Gemini, which is built on the proprietary Gemini models and served only through a restricted API. MedGemma is positioned as a starting point for developers building healthcare applications rather than a finished clinical product.

The initial release spans three configurations: a 4B multimodal model (available as both a pretrained and an instruction-tuned checkpoint), a 27B multimodal instruction-tuned model, and a 27B text-only instruction-tuned model. Image understanding is provided by MedSigLIP, a SigLIP-derived vision encoder pretrained on de-identified medical data spanning chest X-rays, dermatology, ophthalmology, and histopathology. A later MedGemma 1.5 generation, released in January 2026, extends the 4B multimodal model to volumetric CT and MRI, whole-slide histopathology, longitudinal image comparison, bounding-box anatomical localization, medical document understanding, and electronic health record interpretation.

The work was led by Google's health AI teams across Google Research and Google DeepMind. By releasing open weights for a medically tuned multimodal foundation model, MedGemma targets the gap between capable but closed general-purpose systems and the reproducible, adaptable models that biomedical developers need.

#Key Features

• Open, downloadable weights: Unlike API-only medical systems, MedGemma checkpoints are published on Hugging Face, enabling local fine-tuning and private deployment, subject to the Health AI Developer Foundations terms of use.
• Built on Gemma 3: The models inherit Gemma 3's decoder-only transformer design and long context window of at least 128K tokens, available in efficient 4B and 27B sizes.
• MedSigLIP vision encoder: A 400M-parameter SigLIP encoder tuned on de-identified medical images powers image interpretation and is released separately for classification and retrieval tasks.
• Multiple size and modality variants: Text-only and multimodal, 4B and 27B options let developers trade capacity against compute for their use case.
• MedGemma 1.5 extensions: The 1.5 generation adds volumetric imaging, whole-slide pathology, longitudinal comparison, spatial localization, and EHR reasoning to the 4B multimodal model.

#Technical Details

MedGemma uses the Gemma 3 decoder-only transformer as its language backbone in 4B and 27B parameter sizes, with MedSigLIP supplying visual tokens at 448x448 resolution for the multimodal variants. Training combined public datasets — including MIMIC-CXR, SLAKE, PAD-UFES-20, SCIN, TCGA, CAMELYON, and PMC-OA — with proprietary de-identified data across radiology, pathology, dermatology, ophthalmology, and EHR domains. On chest X-ray finding classification the 27B multimodal model reaches a top-5 macro F1 of 90.0 on MIMIC-CXR versus 71.7 for base Gemma 3 27B, and the 27B text-only model scores 89.8 on MedQA (best-of-5). The technical report cites improvements of 2.6-10% on medical multimodal question answering, 15.5-18.1% on chest X-ray finding classification, and 10.8% on agentic evaluations relative to base Gemma, plus a roughly 50% reduction in EHR retrieval errors after fine-tuning.

#Applications

MedGemma is aimed at healthcare and life-science developers who need an open, adaptable base for medical text and image tasks: drafting radiology findings, classifying and interpreting images across radiology, pathology, dermatology, and ophthalmology, answering medical questions, and reasoning over electronic health records. Google frames the released checkpoints as a starting point that developers must further train, validate, and adapt for any specific intended use; the models are not authorized medical devices and require appropriate regulatory clearance before clinical deployment.

#Impact

MedGemma lowers the barrier to building medical multimodal AI by pairing openly downloadable weights with a permissive Apache-2.0 repository of notebooks and detailed model cards, in contrast to closed medical systems reachable only through APIs, making private on-premises adaptation feasible for institutions handling sensitive clinical data. The principal caveats are that the weights are governed by the Health AI Developer Foundations terms of use — a custom Google license with prohibited-use and health-regulatory restrictions rather than an OSI-approved open source license — that the proprietary portion of the training data is not released, and that Google cautions MedGemma is not a clinical device and requires task-specific validation before any real-world medical use.