Foundation model Gated weights (licence click-through) Research use only clinical-LLM

MedGemma 4B multimodal / 27B text

Open-weight medical vision-language models built on Gemma 3: the 4B variant reads chest X-rays, dermatology, ophthalmology and histopathology images alongside text; the 27B variant targets medical text reasoning. Meant as a starting point for developers to fine-tune, not as a finished clinical product.

🤗 google/medgemma-4b-it Paper Docs { } JSON

🤗 1 105 085 downloads / month ♥ 1048 updated 2025-10-28 gated

At a glance

DeveloperGoogle (Health AI Developer Foundations)

Version4B multimodal / 27B text

Released2025-05-20

LicenceHealth AI Developer Foundations terms of use

AvailabilityGated weights (licence click-through)

KindFoundation model

Regulatory statusResearch use only

Parameters4B (multimodal) / 27B (text)

What it does

Not yet documented on this card.

Tasks, data types and cancers

TaskQuestion answering Report generation Classification Information extraction

Data typeMultimodal X-ray Histopathology (WSI)Dermoscopy Clinical text

Clinical settingBasic research Education

CancerPan-cancer

Inputimages + text (instruction format)

Outputtext

Architecture

FamilyGemma 3 decoder LLM + SigLIP image encoder (MedSigLIP)

Parameters4B (multimodal) / 27B (text)

Pre-trainingGemma 3 pretraining + medical image–text and text tuning

Training data

Gemma 3 base plus medical fine-tuning on de-identified image–text data (chest X-ray, dermatology, ophthalmology, histopathology) and medical text; details in the technical report.

Linked datasets

benchmark MedQA (USMLE) Open download

Evaluation

Benchmark / dataset	Metric	Value	External validation	Source
MedQA (USMLE-style, text) ↗	accuracy	reported in technical report (27B variant)	yes	Source

How to run

# generic transformers loader — see the model card for the task-specific head and preprocessing
from transformers import AutoModel, AutoProcessor
model = AutoModel.from_pretrained('google/medgemma-4b-it')
processor = AutoProcessor.from_pretrained('google/medgemma-4b-it')

Librarytransformers

Hardware4B runs on a single 16–24 GB GPU; 27B needs multi-GPU or quantisation

Accept the HAI-DEF terms on Hugging Face. Outputs must be validated per task; Google states the models are not intended for direct clinical use without further evaluation.

Regulatory status and intended use

Regulatory statusResearch use only

Intended useDeveloper foundation for building and validating your own medical applications; not a medical device.

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

Generative: can produce fluent but wrong findings; needs task-specific evaluation and guardrails.
Oncology-specific validation is limited in the report; pathology capability is preliminary.

Sources

At a glance What it does Tasks, data types and cancers Architecture Training data Evaluation How to run Regulatory status and intended use Limitations and bias Sources

Card provenance

Confidencemedium

Verified2026-09-05 · editorial

Updated2026-09-09

Hugging Facesnapshot from 2026-09-09

Facts on this card come only from the listed sources; the Hugging Face figures are a cached public snapshot.

Related models

This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.