Google drops open-source MedGemma models for health AI

MedGemma and MedSigLIP are Google's latest open-weight models built specifically for healthcare AI development, giving you solid foundation models to build on without sending sensitive patient data to third-party APIs.

The release introduces MedGemma 27B Multimodal and a 400M-parameter image encoder called MedSigLIP. MedGemma 27B Multimodal complements the existing 4B multimodal and 27B text-only variants, adding support for complex medical images and longitudinal electronic health record (EHR) interpretation. MedSigLIP handles structured tasks like image classification and embedding-based search.

What it can do and how it works

The MedGemma lineup comes in 4B and 27B sizes, and both now accept images and text to generate text outputs. They are built for free-text tasks like drafting radiology reports, answering visual clinical questions, or summarizing pre-visit patient intakes.

Hardware requirements are refreshingly accessible. Google notes that every model in the release runs on a single GPU. Even better, the 4B MedGemma and the 400M MedSigLIP can be adapted to run directly on mobile hardware and edge devices.

Under the hood, Google trained MedSigLIP on diverse medical imaging data, including chest X-rays, histopathology patches, dermatology images, and fundus photos. They then fine-tuned Gemma 3 models on medical data while intentionally preserving general language skills, instruction-following, and non-English capabilities.

Performance holds up well. With fine-tuning, MedGemma 4B achieves state-of-the-art results on chest X-ray report generation with a RadGraph F1 score of 30.3. Meanwhile, MedSigLIP matches or beats domain-specific vision embedding models while remaining versatile across modalities.

Early adopters are already testing them in the field. DeepHealth in the US is exploring MedSigLIP for chest X-ray triage and nodule detection. Chang Gung Memorial Hospital in Taiwan noted strong results with traditional Chinese medical literature. In India, Tap Health uses MedGemma to summarize progress notes and surface clinical guideline nudges.

Where to get it and current limits

The weights are available now on Hugging Face in safetensors format. Google also published GitHub notebooks covering inference and fine-tuning, alongside dedicated endpoint deployment examples for Vertex AI.

Keep in mind that these are developer foundations, not turnkey medical devices. Google explicitly notes they are not cleared for direct clinical diagnosis, patient management, or standalone treatment decisions without rigorous downstream validation and adaptation.

Why it matters

In healthcare AI, privacy and data sovereignty are dealbreakers. Proprietary cloud APIs are often off-limits due to compliance rules, but training frontier multimodal models from scratch is out of reach for most research teams.

Google made the right tradeoff here: lightweight footprints that fit on a single GPU, paired with open weights developers can actually audit and fine-tune locally. If you are building imaging search, intake automation, or clinical note assistants, this collection should be right at the top of your test list.