Accelerating AI development in healthcare
AI building blocks for creating next-generation healthcare solutions.
HAI-DEF (Health AI Developer Foundations) offers free open-weight models and open-source companion tools with an enabling license to help developers build AI-enabled healthcare applications.
Pre-trained open-weight models
Access pre-trained models to build AI applications, with less data and compute resources.
Customizable for most applications
HAI-DEF model weights are open, so you can run them wherever you'd like and fine-tune them with your data for your tasks.
Integrated with Google Cloud
We make it easy to deploy HAI-DEF models on GCP as scalable services with native integration to Vertex AI and Cloud Healthcare API.
Collection of Gemma 3 variants that are trained for performance on medical text and image comprehension. Developers can use MedGemma to accelerate building healthcare-based AI applications. Can be used for:
- medical image interpretation
- medical text comprehension and clinical reasoning
- privacy-preserving tools for the 3 above use cases within a larger agentic system
MedASR
Lightweight automated speech recognition model with a focus on performance on medical audio, including dictation and notes. Available in English only. Fine-tunable to add new vocabularies and improve performance on specific speakers.
- Dictation across medical specialties for various report and note types
MedSigLIP
Accelerator for medical image analysis. Fine-tuned version of SigLIP for medical images, including chest X-rays, CT slices, MRI slices, dermatology images, ophthalmology images, and histopathology patches. Can be used for:
- Data-efficient classification
- Zero-shot classification
- Semantic image retrieval
TxGemma
Collection of open models to accelerate the development of therapeutics. TxGemma can make predictions about small molecules, proteins, nucleic acids, diseases, and cell lines. Can be used:
- for therapeutic classification, regression, and generation tasks
- conversationally through the chat variants
- for agentic orchestration
HeAR (Lung Acoustics)
Foundation model for non-speech audio analysis. The model produces embeddings that capture dense features relevant for audio applications.
- Data-efficient classification
- Data-efficient regression
Path Foundation
Accelerator for histopathology image analysis. The model produces embeddings that capture dense features relevant for histopathology applications.
- Data-efficient classification
- Similar image search
Legacy Models
The following embeddings models are legacy, but remain available for existing applications. Use MedSigLIP for new development instead of these models:
CXR Foundation
Derm Foundation
