Source: Ollama
Vision / Multimodal Models
MedGemma is Google DeepMind's open-weight model purpose-built for healthcare. It combines a vision encoder trained on medical imaging datasets with a language backbone fine-tuned on clinical literature — so it understands radiology scans, pathology slides, and patient notes in the same model. Deployed privately via Ollama, no protected health information ever leaves your server.
Start Your 90-Day Private Pilot →Analyze chest X-rays, CT slices, dermatology photos, and pathology slides. MedGemma's vision encoder is pre-trained on diverse biomedical imaging corpora — it recognizes anatomical structures and common findings out of the box, without extra fine-tuning for basic use cases.
Summarize discharge summaries, extract ICD-relevant findings from clinical notes, answer questions about lab reports, and draft structured clinical documentation. MedGemma understands medical abbreviations, drug names, and diagnostic terminology that general-purpose models struggle with.
Combine image and text in a single prompt. Ask "Given this chest X-ray and the patient's reported symptoms, what findings are present and what differentials should be considered?" — and receive a structured, medically grounded response entirely within your private environment.
Who this model family is for
MedGemma is a specialized medical model, not a general-purpose chatbot. It can be useful for medical students, residents, or healthcare professionals who want to privately explore clinical reasoning, review sample imaging, or study medical terminology offline. For everyday tasks or general conversation, a standard model like Llama 3.3 or Qwen 2.5 will serve you much better.
Clinics, telehealth providers, and health-tech teams can use MedGemma to build private, on-premise tools for tasks like drafting radiology report summaries, triaging imaging studies, or extracting structured information from clinical notes. Because it runs entirely on your own hardware, sensitive patient data never leaves your infrastructure. It is a decision-support and productivity tool — not a diagnostic system — so all outputs that could affect patient care must be reviewed by a qualified clinician.
Developers building private healthcare applications get a multimodal foundation that understands both medical images (X-rays, pathology, dermatology, ophthalmology) and clinical text. It is well suited for prototyping report generation tools, image analysis assistants, or document parsing pipelines without sending protected health information to external APIs. Expect to invest in prompt engineering, validation, and safety guardrails — MedGemma is a capable building block, not a finished clinical product.
Research teams in radiology, pathology, and clinical NLP can use MedGemma as a strong, openly available baseline for benchmarking and fine-tuning on domain-specific datasets. Its multimodal design makes it valuable for experiments that combine imaging and text. Local deployment allows work with sensitive cohorts under your own data governance rules. All results should be treated as research-grade and validated rigorously against ground truth before any clinical conclusions are drawn.
Licensing information
MedGemma is released by Google under the Health AI Developer Foundations (HAI-DEF) terms, which build on the Gemma license. Commercial use is permitted, but the license includes important health-specific restrictions and responsible-use requirements.
MedGemma is intended as a research and development foundation. It is not a cleared or approved medical device, and its outputs must not be used for direct clinical decision-making without independent validation and appropriate professional oversight.
Official model card:
Disclaimer: NoCloudGPT provides this licensing summary as a convenience reference only. It does not constitute legal advice. License terms can change. Always review the official model card and consult your legal counsel before deploying any model in a commercial or regulated (especially clinical) context.
Available model sizes
| Model Size | Ollama Tag | Recommended NoCloudGPT Tier | Notes | |
|---|---|---|---|---|
| MedGemma 4B (Multimodal) | medgemma:4b |
Starter / Workstation | Handles both medical images and text. Best balance of capability and modest hardware requirements. Good for prototyping. | Get quote → |
| MedGemma 27B (Text) | medgemma:27b |
Professional / Server | Text-only model optimized for advanced clinical reasoning and document-heavy workflows. | Get quote → |
| MedGemma 27B (Multimodal) | medgemma:27b-mm |
Professional / Server | Full image + text understanding at higher capability. Best option for imaging-focused deployments. | Get quote → |
Important note: MedGemma is a specialized medical multimodal model. It should be used with appropriate caution in any clinical or patient-facing context. All outputs that could influence care decisions require review by qualified medical professionals.
Radiology departments running second-read triage tools need a model that never exfiltrates patient data. A MedGemma instance on a dedicated Lightsail 8 GB server — behind your VPN — gives clinicians an AI assistant for preliminary image review, flagging studies that warrant urgent radiologist attention. PHI stays on-prem or in your private cloud account.
CROs handling imaging data from multi-site trials need consistent, auditable analysis without routing scans through consumer AI APIs. MedGemma deployed on a dedicated instance can batch-process trial imaging data, generate structured observation reports, and maintain a complete local audit log — all without a data-sharing agreement headache.
Building a dermatology triage app? An AI ophthalmology screener? MedGemma gives you a medically capable foundation model you can fine-tune on your own labeled dataset and deploy to a server your legal team actually controls. Stop paying per-API-call for a model trained on unknown data with opaque licensing.
Prior authorization teams reviewing submitted clinical documentation and imaging can use MedGemma to extract structured findings, compare against clinical criteria, and pre-populate review forms — processing member records without routing them through a third-party AI vendor.
| Variant | Parameters | Min RAM | Recommended Deployment | Best For |
|---|---|---|---|---|
| MedGemma 4B (text) | 4B | 4 GB | Lightsail 4 GB | Clinical text summarization, note Q&A |
| MedGemma 4B (multimodal) | 4B | 6 GB | Lightsail 8 GB | Image + text, radiology triage |
| MedGemma 27B (multimodal) | 27B | 20 GB+ | GPU instance / on-premise | High-accuracy research, complex multimodal reasoning |
For most clinical workflow pilots, MedGemma 4B multimodal on Lightsail 8 GB is the right starting point. It runs at Q4 quantization (~5–6 GB RAM) without GPU acceleration and handles real-time image + text queries at acceptable clinical-support latency. The 27B variant requires a GPU-equipped instance (A10G or better) for production throughput.
Because MedGemma runs entirely on your infrastructure via Ollama, patient images and clinical notes never leave your environment. There's no API call to Google, no model-provider data retention, and no ambiguity about where your data goes.
AWS Lightsail instances can be deployed in HIPAA-eligible regions under your existing AWS Business Associate Agreement. Combine with encrypted EBS volumes and VPC isolation for a defensible HIPAA technical safeguard posture — without standing up a complex managed AI service.
Every query and response can be logged locally to structured JSON. Your compliance team gets a complete audit trail of AI-assisted clinical decisions — query, image hash, response, timestamp — without exporting sensitive data to a vendor's logging infrastructure.
MedGemma's open weights mean you can fine-tune the model on your institution's labeled imaging dataset using LoRA adapters — then deploy the adapted model privately. Your institution's clinical knowledge stays yours, not incorporated into a shared vendor model.
MedGemma is a research and clinical decision support tool. It is not FDA-cleared or CE-marked as a medical device and should not be used as the sole basis for diagnostic or treatment decisions. All AI-generated clinical outputs must be reviewed by qualified healthcare professionals. Confirm your specific regulatory obligations with your compliance and legal teams before deployment.
curl -fsSL https://ollama.com/install.sh | sh
ollama pull medgemma
http://localhost:11434 via Ollama's REST API.
Run medical-grade AI on your infrastructure. NoCloudGPT helps you deploy MedGemma with a HIPAA-aligned configuration, Ollama setup, and 90-day pilot plan — so your clinical team can evaluate it before committing to any long-term infrastructure spend.
Artwork and source references
Hero image
Developer reference
MedGemma is developed by Google DeepMind. NoCloudGPT is an independent deployment service not affiliated with Google DeepMind.
Model catalog
See Ollama library: medgemma for current tags and sizes.