Quick models guide

Model Card	Open Weight	Research Endpoint*	Demos	Shortcuts
MedGemma	Yes (Hugging Face, Model Garden)	No		More Shortcuts Common Use Cases: Common Use Cases Adapting MedGemma: Adapting MedGemma
CXR Foundation	Yes (Hugging Face, Model Garden)	Yes (Sign-up Form)		More Shortcuts Common Use Cases: Developing data-efficient classifiers Zero-shot classification Semantic image retrieval Developer Resources: Docs GitHub If you'd like to access the model using a research endpoint instead, please fill out the access request form. Note that CXR Foundation uses multiple underlying models to process your requests. The open source Serving API handles the model orchestration for you; whereas with the research endpoint, you coordinate the use of the models as the caller. Datasets: NCI Imaging Data Commons (IDC) NIH Chest X-ray Encoded NIH subset Publications: Blog post 07/22 Blog post 08/23 Blog post 11/24 Preprint
CT Foundation	No	Yes (Sign-up Form)		More Shortcuts Common Use Case: Developing data-efficient classifiers Developer Resources: GitHub The model is only accessible as a research endpoint. Fill out the access request form to gain access. Once you have access to the research endpoint, use the demo notebook as your starting point. Datasets: National Cancer Institute - Imaging Data Commons public datasets (IDC): learn; explore; request access and use it on GCP. Alternatively, use the pre-encoded subset. LIDC: learn and request access to use it in Cloud DICOM Store projects/hai-cd3-foundations/locations/us-central1/datasets/ct3d. NLST: learn; request access and use the pre-encoded subset. Publications: blog post 10/24, preprint
Derm Foundation	Yes (Hugging Face, Model Garden)	Yes (Sign-up Form)		More Shortcuts Common Use Case: developing data-efficient classifiers Developer Resources: Docs GitHub If you'd like to access the model using a research endpoint instead, please fill out the access request form. Datasets: Skin Condition Image Network (SCIN): learn; use the images; or pre-encoded. PAD-UFES-20: learn; use the images; or pre-encoded. DERM12345 Embeddings and Demo: pre-encoded embeddings and data available for DERM12345). Special thanks to Abdurrahim Yilmaz for providing this. Publications: Blog post 03/24 Blog post 11/24
HeAR (acoustic)	Yes (Hugging Face, Model Garden)	Yes (Sign-up Form)		More Shortcuts Common Use Case: developing data-efficient classifiers Developer Resources: Docs GitHub If you'd like to access the model using a research endpoint instead, please fill out the access request form to gain access. Once you have access to the research endpoint, use the demo notebook as your starting point. Datasets: CIDRZ Publications: Blog post 08/24 Preprint
Med-Gemini	No	Program closed (Sign up for updates)		More Shortcuts The trusted tester program is now closed, but you can sign up to receive updates. To learn more about the model, refer to the blog post or the preprint.
Path Foundation	Yes (Hugging Face, Model Garden)	Yes (Sign-up Form)		More Shortcuts Common Use Cases: Developing data-efficient classifiers Similar image search Developer Resources: Docs GitHub If you'd like to access the model using a research endpoint instead, please fill out the access request form. Datasets: National Cancer Institute - Imaging Data Commons public datasets (IDC): learn; explore; request access and use it on GCP. Alternatively, use the pre-encoded subset. Camelyon17: learn; use in on GCS; or request access and use it on Cloud DICOM Store projects/hai-cd3-foundations/locations/us-west1/datasets/pathology/dicomStores/camelyon. Publications: Blog post 03/24 Blog post 11/24 Preprint
TxGemma	Yes (Hugging Face, Model Garden)	No		More Shortcuts Common Use Cases: Therapeutic classification, regression, and generation tasks Agentic orchestration Developer Resources: Docs GitHub Publications: Blog post Paper

*Research Endpoints: In addition to the publicly available resources on Hugging Face and Model Garden, HAI-DEF provides online service endpoints for many of its externalized models, offering a research-use only pathway, often requiring less setup and infrastructure for researchers looking to quickly experiment with and evaluate HAI-DEF's models; read more.