MedGemma

Google's open-weight Gemma 3-based model family for medical text and multimodal clinical reasoning (dermatology, pathology, radiology, clinical text).

Unsigned

Overview

MedGemma is Google's open-weight Gemma 3-based model collection for medical text and multimodal medical image comprehension, covering dermatology, pathology, ophthalmology, chest X-ray, and clinical text, released as 4B multimodal and 27B text/multimodal variants. It is trained and evaluated on datasets including MIMIC-CXR, TCGA (cancer genomics data), CAMELYON histopathology, PAD-UFES-20, and SLAKE, and is documented with an unusually formal, separately maintained structured model-card document (developers.google.com/health-ai-developer-foundations/medgemma/model-card) covering intended use, training data, and evaluation, positioning it for clinical-note extraction and clinical-image interpretation workflows that match the PRD's own named example.

**Marginal call, stated plainly**: MedGemma is a general clinical/health model (dermatology, radiology, pathology, clinical text), not a genomics-native model, even though its training data includes TCGA cancer genomics data as one dataset among many. It implements no GA4GH standard or extension. It is catalogued here as a marginal fit on the genomics-specificity axis, not a clean one, consistent with how this registry has already flagged other general clinical/health models.

**Licence note, confirmed directly from Google's own model-card documentation** (not derived Hugging Face metadata alone): "The use of MedGemma is governed by the Health AI Developer Foundations terms of use" -- a custom, gated licence (Hugging Face access requires accepting these terms), not a standard open-source SPDX licence, hence `license: other` with `license_url` pointing to the terms document itself.

**Registry note on `model_card_uri`**: this points to Google's own separately published, structured MedGemma model-card document rather than the gated Hugging Face repository page, since the former is the more authoritative and more fully documented source.

Details

Licence: other (full text)
Version: MedGemma 1.5 (4B multimodal instruction-tuned)
Category: Multimodal genomic or health model
Homepage: https://developers.google.com/health-ai-developer-foundations/medgemma
Repository: https://github.com/google-health/medgemma
Maintainers: Susheel Varma (@susheel, GA4GH AI Workstream / Sage Bionetworks)
Keywords: clinical-textmultimodalmedical-imaginggemmahealth-ai

Safety classification

No safety classification has been submitted for this entry.

Model Card

View the full GA4GH Genomic AI Model Card ↗

Hosting: huggingface
Access: https://huggingface.co/google/medgemma-4b-it

Mirrored data last synced from the Model Card at 2026-09-08T00:00:00.000Z. See the linked card for the authoritative, live version.

Provenance

Created: 2026-09-08T00:00:00.000Z
Updated: 2026-09-08T00:00:00.000Z
Last verified: 2026-09-08T00:00:00.000Z