Healthcare is increasingly using AI to improve workflow management, patient communication, and diagnostic and treatment support, and it's critical that these AI-based systems are high-performing, efficient, and privacy-preserving. To address this, Health AI Developer Foundations (HAI-DEF) was built, a collection of lightweight open models that offer developers robust starting points for their own health research and application development. HAI-DEF models are open, allowing developers to retain full control over privacy, infrastructure, and modifications to the models. The MedGemma collection, part of HAI-DEF, includes variants in 4B and 27B sizes that accept image and text inputs and produce text outputs. MedGemma models are strong starting points for medical research and product development, useful for medical text or imaging tasks that require generating free text. MedSigLIP is a lightweight image and text encoder for classification, search, and related tasks, and is recommended for imaging tasks that involve structured outputs like classification or retrieval. All MedGemma and MedSigLIP models can be run on a single GPU, and some can even be adapted to run on mobile hardware. The MedGemma collection is open, allowing developers to download, build upon, and fine-tune the models to support their specific needs. Researchers and developers have been exploring the MedGemma models for their use cases, finding them adept at solving crucial problems. To help developers get started, detailed notebooks on GitHub demonstrate how to create instances of MedSigLIP and MedGemma for both inference and fine-tuning on Hugging Face.