Medical AI Background

MedGemma

Advanced AI Models for Medical Text and Image Analysis

Powering the next generation of healthcare applications with Google DeepMind's cutting-edge MedGemma AI models for medical understanding.

2
Model Variants
4B
Multimodal Model
27B
Text-Only Model

Try MedGemma Interactive Demo

Experience the power of MedGemma 4B IT model for medical text and image analysis

What is MedGemma?

MedGemma is a collection of cutting-edge AI models designed specifically to understand and process medical text and images. Developed by Google DeepMind and announced in May 2025, MedGemma represents a significant advancement in the field of medical artificial intelligence.

Built on the powerful Gemma 3 architecture, MedGemma has been optimized for healthcare applications, providing developers with robust tools to create innovative medical solutions.

As part of the Health AI Developer Foundations, MedGemma aims to democratize access to advanced medical AI technology, enabling researchers and developers worldwide to build more effective healthcare applications.

Recent Development

Launched at Google I/O 2025

May
2025

Released as part of Google's ongoing efforts to enhance healthcare through technology

Features of MedGemma

Powerful capabilities designed for medical applications

MedGemma Model Variants

4B Multimodal Model

Processes both medical images and text with 4 billion parameters, using a SigLIP image encoder pre-trained on de-identified medical data.

27B Text-Only Model

Optimized for deep medical text comprehension and clinical reasoning with 27 billion parameters.

Key Capabilities

  • Medical image classification (radiology, pathology, etc.)
  • Medical image interpretation and report generation
  • Medical text comprehension and clinical reasoning
  • Patient preclinical interviews and triaging
  • Clinical decision support and summarization

Performance Comparison

Use Cases for MedGemma

Healthcare Application Development

Build AI-based applications that examine medical images, generate reports, and triage patients.

Medical Research and Innovation

Accelerate research with open access to advanced AI through Hugging Face and Google Cloud.

Clinical Support Roles

Enhance patient interviewing and clinical decision support for improved healthcare efficiency.

How to Use MedGemma

Implementation guides and adaptation methods

1

Access MedGemma Models

MedGemma models are accessible on platforms like Hugging Face, subject to the terms of use by the Health AI Developer Foundations.

# Example Python code to load MedGemma model
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("google/medgemma-4b-it")
model = AutoModelForCausalLM.from_pretrained("google/medgemma-4b-it")
2

Adaptation Methods

Prompt Engineering

Use few-shot examples and break tasks into subtasks to enhance performance.

Fine-Tuning

Optimize using your own medical data with resources like GitHub notebooks.

Agentic Orchestration

Integrate with tools like web search, FHIR generators, and Gemini Live.

3

Deployment Options

Choose the right deployment method based on your requirements:

Local Deployment

Run models locally for experimentation and development purposes.

Cloud Deployment

Deploy as scalable HTTPS endpoints on Vertex AI through Model Garden for production-grade applications.

Implementation Considerations

Validation Requirements

MedGemma models are not clinical-grade out of the box. Developers must validate performance and make necessary improvements before deploying in production environments.

Terms of Use

The use of MedGemma is governed by the Health AI Developer Foundations terms of use, which developers must review and agree to before accessing models.

Frequently Asked Questions

Common questions about MedGemma

What are the key differences between the 4B multimodal and 27B text-only MedGemma models?

The 4B multimodal model processes both medical images and text with 4 billion parameters, using a SigLIP image encoder. The 27B text-only model focuses exclusively on text processing with 27 billion parameters, optimized for deeper medical text comprehension and clinical reasoning.

Are MedGemma models ready for clinical use out of the box?

No, MedGemma models are not considered clinical-grade out of the box. Developers must validate their performance and make necessary improvements before deploying in production environments, especially for applications involving patient care.

How can I access MedGemma models for my development work?

MedGemma models are accessible on platforms like Hugging Face and Google Cloud, subject to the terms of use by the Health AI Developer Foundations. You can run them locally for experimentation or deploy them via Google Cloud for production-grade applications.

What types of medical images can the 4B multimodal model process?

The 4B multimodal model is pre-trained on diverse medical images including chest X-rays, dermatology images, ophthalmology images, and histopathology slides, making it adaptable for various medical imaging tasks.

What adaptation methods can improve MedGemma's performance for specific tasks?

Developers can use prompt engineering (few-shot examples), fine-tuning with their own medical data, and agentic orchestration with tools like web search, FHIR generators, and Gemini Live to enhance performance for specific use cases.

When was MedGemma released and by whom?

MedGemma was officially launched around May 20-22, 2025, during Google I/O 2025 by Google DeepMind, as part of their ongoing efforts to enhance healthcare through technology.

How does MedGemma compare to similar models of its size?

According to its model card on Google Developers, MedGemma's baseline performance is strong compared to similar-sized models. It has been evaluated on clinically relevant benchmarks, including open datasets and curated datasets, with a focus on expert human evaluations for tasks.

Are there any resources available for fine-tuning MedGemma?

Yes, resources including notebooks on GitHub are available to facilitate fine-tuning, such as a fine-tuning example using LoRA available at Google's MedGemma GitHub repository.

What are the hardware requirements for running MedGemma models?

The hardware requirements depend on the model variant. According to posts from Google AI, MedGemma models are designed to be efficient, with the ability to run fine-tuning and inference on a single GPU, making them more accessible than some larger models.

Does MedGemma support multilingual medical terminology?

Based on community discussions, there are questions about MedGemma's performance with non-English medical terminology, such as Japanese medical terms. This suggests that multilingual support may vary and could be an area for future improvement or fine-tuning.