In the modern medical field, artificial intelligence (AI) is gradually becoming an important tool for improving work efficiency, enhancing patient communication, and supporting diagnosis and treatment. To meet the demands of medical AI systems in terms of performance, efficiency, and privacy protection, Health AI Developer Foundations (HAI-DEF) was born.
This project includes a series of lightweight open models, designed to provide developers with a strong foundation to support their health research and application development. The openness of HAI-DEF models ensures that developers can fully control data privacy, infrastructure, and model modifications.
In May of this year, we expanded HAI-DEF and launched MedGemma, a set of generative models based on Gemma3, aimed at accelerating AI development in the medical and life sciences fields. Recently, we also released two new models: the MedGemma27B multimodal model and MedSigLIP. The MedGemma27B multimodal model adds support for interpreting complex multimodal and longitudinal electronic health records, building upon the existing 4B multimodal and 27B text models. MedSigLIP is a lightweight image and text encoder suitable for tasks such as classification and search.
The MedGemma and MedSigLIP models provide a good starting point for medical research and product development. MedGemma is particularly suitable for medical tasks requiring the generation of free-form text, such as report generation or visual question answering; while MedSigLIP is recommended for imaging tasks that require structured outputs, such as classification or retrieval. These models can run on a single GPU, and MedGemma4B and MedSigLIP can also be adapted to mobile hardware.
The openness of the MedGemma series models allows developers to download, build, and fine-tune them according to specific needs. This open approach offers significant advantages in the medical field compared to API models. Developers can run the models in their preferred environment, flexibly addressing privacy issues and institutional policies; at the same time, by fine-tuning and modifying, developers can optimize model performance, ensuring stability and reproducibility, which is particularly important in medical applications.
To help developers get started quickly, we have provided detailed notebooks on GitHub, demonstrating how to create instances of MedSigLIP and MedGemma on the Hugging Face platform and perform inference and fine-tuning. In addition, MedGemma and MedSigLIP can be seamlessly deployed into Vertex AI, providing dedicated endpoint support.
Blog: https://research.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/
Key Points:
🌟 HAI-DEF introduces MedGemma and MedSigLIP, providing strong support for health AI development.
🔍 MedGemma is suitable for generating free-form text tasks, while MedSigLIP focuses on image classification and retrieval.
🔑 The flexibility of open models allows developers to safely optimize and fine-tune models in local environments.