We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
DeepMind Blog·Oct 25·5 min read
MedGemma: Our most capable open models for health AI development
TL;DR
Healthcare AI systems face challenges in performance, efficiency, and privacy. DeepMind introduced MedGemma, a collection of advanced open models for health AI, including MedGemma 27B Multimodal and MedSigLIP.
✦ Why It Matters
Engineers can leverage MedGemma models to build efficient, privacy-conscious health AI applications.
Key Takeaways
- Utilize MedGemma for advanced healthcare AI applications.
- Leverage MedSigLIP for efficient image classification tasks.
- Run models on single GPUs or adapt for mobile use.
- Access full technical details in the MedGemma report.
Full Summary
MedGemma is part of DeepMind's Health AI Developer Foundations (HAI-DEF), which provides lightweight open models for healthcare AI development. The new MedGemma models include a 27B Multimodal variant that interprets complex electronic health records and a lightweight image and text encoder called MedSigLIP.
These models support tasks such as report generation and image classification, making them versatile tools for medical research. MedGemma 4B scored 64.4% on the MedQA benchmark, indicating strong performance for smaller models.
All models can run on a single GPU, with some capable of mobile adaptation, making them accessible for various applications. The release aims to empower developers with control over privacy and infrastructure in health AI projects.