Powered by Dr7.ai Medical AI Platform
MedGemma is Google's family of open-weight medical AI models built on Gemma 3. Explore versions, benchmarks and downloads, or try the free MedGemma interactive demo below.
Experience the power of Medical Gemma 4B IT model for advanced medical text and image analysis
Experience AI Medical Assistant
Use Dr7's own medical AI assistant through an OpenAI-compatible API. No GPU or deployment to manage.
Upgrade to Pro version to unlock advanced AI models, unlimited conversations, professional medical image analysis and more powerful features
MedGemma is a family of open AI models from Google, built on Gemma 3 and trained to understand medical text and medical images. Google released the first MedGemma models on 20 May 2025; the newest release is MedGemma 1.5 4B (13 January 2026).
Built on the powerful Gemma 3 architecture, MedGemma has been optimized for healthcare applications, providing developers with robust tools to create innovative medical solutions using MedGemma's advanced capabilities.
As part of the Health AI Developer Foundations, MedGemma aims to democratize access to advanced medical AI technology, enabling researchers and developers worldwide to build more effective healthcare applications with MedGemma. You may also see it written as Med Gemma, Med-Gemma, MedGamma, Medgema or Google MedGemma — these all refer to the same model family. MedGemma is not the same as Med-Gemini, a separate Google research model that has not been released publicly.
Adds 3D CT and MRI volumes, whole-slide pathology, comparison with prior chest X-rays, anatomy localization and lab-report data extraction.
MedGemma 1.5 4B in detailChecked October 2026: there is no MedGemma 2 and no MedGemma 1.5 27B — MedGemma 1.5 exists only as a 4B model.
Every official MedGemma checkpoint published by Google, plus the related MedSigLIP image encoder.
Updated 4B model; only available as instruction-tuned 4B.
Details: MedGemma 1.5 4BInstruction-tuned — the recommended starting point.
Pre-trained base model for deeper experimentation.
Trained only on medical text; highest text-benchmark scores.
One 27B model for text, medical records and images.
Not a chat model: the medical image encoder behind MedGemma, for classification and image search.
Details: MedSigLIP| Model | Parameters | Input | Released | Hugging Face repo | License |
|---|---|---|---|---|---|
MedGemma 1.5 4BLatest Updated 4B model; only available as instruction-tuned 4B. | 4B | Text + images, incl. 3D CT/MRI and whole-slide pathology | Jan 2026 | google/medgemma-1.5-4b-it | HAI-DEF terms |
MedGemma 4B-it Instruction-tuned — the recommended starting point. | 4B | Text + images | May 2025 | google/medgemma-4b-it | HAI-DEF terms |
MedGemma 4B-pt Pre-trained base model for deeper experimentation. | 4B | Text + images | May 2025 | google/medgemma-4b-pt | HAI-DEF terms |
MedGemma 27B text-it Trained only on medical text; highest text-benchmark scores. | 27B | Text only | May 2025 | google/medgemma-27b-text-it | HAI-DEF terms |
MedGemma 27B multimodal One 27B model for text, medical records and images. | 27B | Text + images + EHR data | Jul 2025 | google/medgemma-27b-it | HAI-DEF terms |
MedSigLIPRelated Not a chat model: the medical image encoder behind MedGemma, for classification and image search. | 400M + 400M | Image and text embeddings (448×448) | Jul 2025 | google/medsiglip-448 | HAI-DEF terms |
Sources: MedGemma 1.5 model card · MedGemma 1 model card · MedGemma Technical Report (arXiv 2507.05201) · MedGemma 1.5 Technical Report (arXiv 2604.05081) · Nature Medicine (2026) · MedSigLIP model card
What MedGemma can do, according to Google's model cards
Read medical images and text together. They use a SigLIP image encoder pre-trained on de-identified medical data — chest X-rays, dermatology, ophthalmology and histopathology images.
Trained exclusively on medical text and optimized for medical reasoning, so it scores slightly higher than the 27B multimodal model on text benchmarks.
Accuracy (%) or macro F1 as reported in Google's official model cards. 27B results use test-time scaling.
| Benchmark | MedGemma 4B | MedGemma 1.5 4B | MedGemma 27B text | MedGemma 27B multimodal |
|---|---|---|---|---|
| MedQA (4 options) | 64.4 | 69.1 | 87.7 | 85.3 |
| MedMCQA | 55.7 | 59.8 | 74.2 | 70.2 |
| PubMedQA | 73.4 | 68.2 | 76.8 | 77.2 |
| MMLU medical | 70.0 | 69.6 | 87.0 | 86.2 |
| MIMIC-CXR (macro F1, top 5 conditions) | 88.9 | 89.5 | — | 90.0 |
| EyePACS fundus | 64.9 | 76.8 | — | 75.3 |
Sources: MedGemma 1.5 model card · MedGemma 1 model card
Build AI-based applications using MedGemma that examine medical images, generate reports, and triage patients with MedGemma's advanced capabilities.
Accelerate research with open access to MedGemma advanced medical AI models through various platforms and cloud services supporting MedGemma.
Prototype decision-support, summarization and record-review tools, then validate them before any clinical use.
Run MedGemma yourself with Hugging Face or Vertex AI, or try Dr7's medical AI chat online. Need the Dr7 API? It's right below the demo.
Ask Dr7's own medical AI assistant a medical question or upload an X-ray, CT slice or skin photo. Sign in to start.
Open the demoSign in to Hugging Face, open the model page and accept the Health AI Developer Foundations terms (access is granted immediately). Then load it with Transformers 4.50 or newer. Google says every MedGemma model can run on a single GPU; the 27B models need a large one unless you quantize them.
MedGemma on Hugging FaceDeploy MedGemma as a managed endpoint from Vertex AI Model Garden in your own Google Cloud project — Google's recommended path for production scale. Billed by Google Cloud.
Vertex AI Model Garden# pip install -U "transformers>=4.50" accelerate torch pillow requests
from transformers import pipeline
from PIL import Image
import requests
import torch
pipe = pipeline(
"image-text-to-text",
model="google/medgemma-4b-it", # or "google/medgemma-1.5-4b-it"
torch_dtype=torch.bfloat16,
device="cuda",
)
# Image attribution: Stillwaterising, CC0, via Wikimedia Commons
image_url = "https://upload.wikimedia.org/wikipedia/commons/c/c8/Chest_Xray_PA_3-8-2010.png"
image = Image.open(requests.get(image_url, headers={"User-Agent": "example"}, stream=True, timeout=30).raw)
messages = [{"role": "user", "content": [
{"type": "image", "image": image},
{"type": "text", "text": "Describe this X-ray"},
]}]
output = pipe(text=messages, max_new_tokens=300)
print(output[0]["generated_text"][-1]["content"])Use few-shot examples and break tasks into smaller subtasks to improve results.
Train on your own medical data, starting from Google's fine-tuning notebook on GitHub.
Combine MedGemma with tools such as web search, FHIR data generators or other AI models.
MedGemma is not clinical-grade out of the box. Google's model card states that developers must validate and adapt it for their specific use case before relying on it.
Use of MedGemma is governed by the Health AI Developer Foundations terms of use, which you accept before downloading the weights.
Which medical AI models can you actually download, what they read, and their published MedQA scores.
| Model | Developer | Downloadable? | Input | MedQA |
|---|---|---|---|---|
| MedGemma 4B | Yes (accept HAI-DEF terms) | Text + images | 64.4% | |
| MedGemma 1.5 4B | Yes (accept HAI-DEF terms) | Text + images (incl. 3D CT/MRI) | 69.1% | |
| MedGemma 27B text | Yes (accept HAI-DEF terms) | Text | 87.7% | |
| MedGemma 27B multimodal | Yes (accept HAI-DEF terms) | Text + images | 85.3% | |
| Med-PaLM 2 | No — weights never released | Text | 86.5% | |
| LLaVA-Med v1.5 | Microsoft | Yes (research use only) | Text + images | Not reported |
Scores come from each model's own paper or model card under different settings (for example, Med-PaLM 2 used ensemble refinement prompting and MedGemma 27B used test-time scaling), so treat them as indicative rather than a head-to-head test. LLaVA-Med was evaluated on medical image question answering (VQA-RAD, SLAKE, PathVQA), not MedQA.
Sources: MedGemma 1.5 model card · MedGemma 1 model card · Med-PaLM 2 (arXiv 2305.09617) · LLaVA-Med (arXiv 2306.00890) · microsoft/llava-med-v1.5-mistral-7b
The models described on this page are research tools for clinical imaging. If what you have is an everyday photo — a rash, a mole, a nail, a wound — AI Doctor Scan reads it and returns a free urgency level and a plain-language summary in about 30 seconds. It does not name a diagnosis, and it needs no account.
Try AI Doctor Scan — freeCommon questions about MedGemma
MedGemma 4B is a 4-billion-parameter multimodal model that reads medical images and text. MedGemma 27B comes in two 27-billion-parameter versions: a text-only model trained exclusively on medical text, and a multimodal model that also handles images and electronic health record data. The 27B models score higher on medical text benchmarks (MedQA 87.7% for 27B text-only vs 64.4% for 4B), while 4B is much cheaper to run. MedGemma 1.5 4B (January 2026) is an updated 4B model.
No. Google's model card says MedGemma is not intended to be used without appropriate validation, adaptation or meaningful modification for a specific use case, and its outputs are not intended to directly inform clinical diagnosis, patient management or treatment decisions.
MedGemma weights are on Hugging Face (for example google/medgemma-4b-it). Sign in, accept the Health AI Developer Foundations terms on the model page — access is granted immediately — and load the model with Transformers. You can also deploy it from Google Cloud's Vertex AI Model Garden.
MedGemma's image encoder was pre-trained on de-identified chest X-rays, dermatology images, ophthalmology (fundus) images and histopathology slides. MedGemma 1.5 4B adds 3D CT and MRI volumes, multiple patches from whole-slide pathology images, and comparison of a chest X-ray with prior images.
Google lists prompt engineering (for example few-shot examples), grounding, agentic orchestration with other tools, and fine-tuning on your own data. The official google-health/medgemma GitHub repository includes a fine-tuning notebook.
Google released MedGemma 4B and MedGemma 27B text-only on 20 May 2025, around Google I/O. MedGemma 27B multimodal and the MedSigLIP image encoder followed on 9 July 2025, and MedGemma 1.5 4B on 13 January 2026. The peer-reviewed MedGemma paper appeared in Nature Medicine on 6 October 2026.
On MedQA (USMLE-style, 4 options), Google reports 64.4% for MedGemma 4B, 69.1% for MedGemma 1.5 4B, 87.7% for MedGemma 27B text-only and 85.3% for MedGemma 27B multimodal (27B results use test-time scaling). For comparison, the base Gemma 3 4B model scores 50.7%.
Yes. Google's google-health/medgemma GitHub repository provides a quick-start notebook and a fine-tuning notebook for Hugging Face, plus further tutorial notebooks, for example for CT and whole-slide pathology with MedGemma 1.5.
Google says all MedGemma models can run on a single GPU, and that MedGemma 4B can even be adapted to run on mobile hardware. The 27B models need a large GPU — Google's quick-start notes point to Colab Enterprise-class GPUs to run them without quantization.
Not as of October 2026. The newest release is MedGemma 1.5, which Google offers only as a 4B multimodal instruction-tuned model. The 27B models are still the original MedGemma 1 versions from May and July 2025.
MedGemma is built on Gemma 3, which supports over 140 languages, so it can read and answer in many languages. However, Google notes that its safety evaluations used primarily English prompts and its medical benchmarks are mostly English, so you should validate quality in your own language (for example Japanese or Spanish) before relying on it.
No. Med-Gemini (2024, 91.1% on MedQA) and Med-PaLM 2 (2023, 86.5% on MedQA) are Google research models whose weights were never released. MedGemma is Google's open model family that anyone can download and fine-tune.