Google's MedGemma: Open Medical AI You Can Run Yourself

Quick answer

MedGemma is a free-to-download family of medical AI models from Google that can read medical text and images (like chest X-rays, skin photos and lab reports) and write text answers. You can run it on your own computer or server, so patient data never has to leave your building. But it is a starting point for developers, not a ready-made doctor: Google's own model card says its outputs "are not intended to directly inform clinical diagnosis" and must be validated first (Google for Developers, MedGemma model card).

1. Sizes: 4B multimodal (text + images), 27B multimodal, 27B text-only, plus a 400M-parameter image encoder called MedSigLIP (Nature Medicine, 6 Oct 2026).
2. Licence: not Apache or MIT. It is governed by Google's Health AI Developer Foundations (HAI-DEF) Terms of Use (Google, last modified 15 Nov 2024).
3. Why it is in the news now: the peer-reviewed MedGemma paper appeared in Nature Medicine on 6 October 2026, and Google says the models have passed 10 million downloads (Google Research blog, 23 Oct 2026).

Honesty note: SR Infobiz has not run MedGemma or tested it ourselves. Every number below comes from the named source, and most benchmark numbers are Google's own. This article is technology news and explanation, not medical advice.

What exactly is MedGemma?

MedGemma is a collection of Google's open-weight models "for medical text and image comprehension, built on Gemma 3" (Google for Developers, MedGemma overview). "Open-weight" means Google publishes the trained model files, so you can download them and run them yourself instead of calling a paid cloud API.

Is this a brand-new release?

Not quite, and it is worth being clear about this. Google's release notes show the first MedGemma 4B model went out on 20 May 2025, with a bug fix on 9 July 2025 and an upgraded MedGemma 1.5 4B on 13 January 2026 (Google, MedGemma model card release notes). What happened in October 2026 is the peer-reviewed paper in Nature Medicine, published 6 October 2026, which news sites such as AI Weekly covered around 17–20 October 2026 (AI Weekly, 20 Oct 2026). On 23 October 2026, Google Research added that MedGemma has "seen more than 10 million downloads worldwide" (Google Research blog, 23 Oct 2026).

The model line-up

Source: Google for Developers model card and Hugging Face model card (accessed 24 Oct 2026); Nature Medicine, 6 Oct 2026.

All variants output text only. Images are resized to 896 × 896 and each takes 256 tokens; the models accept up to 128K tokens of input and produce up to 8,192 tokens of output (Google, MedGemma model card).

How MedGemma works, in plain English

Think of MedGemma as two parts glued together:

1. An "eye" (the image encoder). Google took the image encoder from Gemma 3 (called SigLIP) and retrained it on millions of medical image–text pairs. Medical data was kept to about 2% of that training mix so the encoder would not forget general images (Nature Medicine, 6 Oct 2026). This medical eye is also released separately as MedSigLIP.
2. A "brain" (the language model). This is Gemma 3, further trained with medical images mixed in at about 12% of the data, then taught to follow instructions using distillation and reinforcement learning (Nature Medicine, 6 Oct 2026).

You give it an image and a question ("Describe this X-ray") or just text ("Summarise this discharge note"), and it writes an answer. Training data included chest X-rays, dermatology, eye (fundus) and pathology images, medical Q&A, and synthetic health records, all described by Google as de-identified or licensed (Google, MedGemma model card).

What can MedGemma do? (Google's own numbers)

Every figure in this section was reported by Google in its model card or its Nature Medicine paper. Independent results are labelled separately. We did not verify any of them.

Medical exam-style questions

1. On MedQA (US medical licensing–style questions), the paper reports 86.2 for MedGemma 27B, which the authors call "the best performance for an open model at the time of its release, other than the 671-billion-parameter DeepSeek-R1 model (91.0)" (Nature Medicine, 6 Oct 2026). Google's model card lists 85.3 for MedGemma 1 27B on MedQA (4-option), so expect small differences between documents.
2. The small model improved over its base: MedQA went from 50.7 (Gemma 3 4B) to 64.4 (MedGemma 1 4B) and 69.1 (MedGemma 1.5 4B) (Google, MedGemma 1.5 model card).
3. On MedMCQA, a dataset built by Saama AI Research in Chennai from Indian medical entrance-exam questions, MedGemma 27B scored 70.2 and MedGemma 1.5 4B scored 59.8 (Google, MedGemma 1.5 model card).

Reading medical images

1. On chest X-ray classification (MIMIC-CXR, top 5 conditions, macro F1), Gemma 3 4B scored 81.2 and MedGemma 1.5 4B scored 89.5 (Google, MedGemma 1.5 model card).
2. On diabetic-retinopathy eye images (EyePACS, an internal dataset), accuracy rose from 14.4 for Gemma 3 4B to 76.8 for MedGemma 1.5 4B (Google, MedGemma 1.5 model card).
3. On data it was not trained on, the paper reports gains over the base model of 2.6–10% in medical image question answering and 15.5–18.1% in chest X-ray finding classification (Nature Medicine, 6 Oct 2026).

Writing X-ray reports

The headline claim: 81% of MedGemma-generated chest X-ray reports were rated "as likely to result in the same or superior clinical decisions" compared with the original reports (Nature Medicine, 6 Oct 2026). Important context from the same paper: this was judged by a single US board-certified thoracic radiologist who knew which reports were AI-generated. That is a useful signal, not proof of clinical safety.

Lab reports and health records (MedGemma 1.5)

Google says MedGemma 1.5 4B turns lab-report PDFs into structured JSON with a macro F1 of 91.0 on one internal dataset, versus 84.0 for Gemma 3 4B, and answers questions on synthetic health records (EHRQA) with 89.6% accuracy (Google, MedGemma 1.5 model card). For Indian health-tech teams digitising paper lab reports, this is probably the most practical feature.

What independent testers found

The paper cites third-party results: on the public ReXrank chest X-ray leaderboard, MedGemma 4B had the top GREEN score on two datasets but was "middle-of-the-leaderboard" on others (Nature Medicine, 6 Oct 2026, citing ReXrank.ai, accessed 9 Feb 2026). On the MedHallu hard set, which tests whether a model can spot a plausible-but-wrong answer, MedGemma 4B scored 0.50 versus 0.09 for Gemma 3 4B (Nature Medicine, 6 Oct 2026, citing the Medmarks study). In short: strong for its size, not best at everything.

What MedGemma can't do (limits and safety)

Google is unusually direct about the limits. From its model card (accessed 24 Oct 2026):

1. Not for direct clinical use. "The outputs generated by MedGemma are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications."
2. Not clinical-grade. Google's overview page says MedGemma "isn't yet clinical-grade and will likely require further fine-tuning."
3. Mistakes happen. "Even for image and text domains that constitute a substantial portion of training data, inaccurate model output is possible."
4. Narrow testing. It was mainly evaluated on single-image tasks and "has not been evaluated or optimized for multi-turn applications" (MedGemma 1 card). Safety testing used "primarily English language prompts", which matters for Hindi, Tamil, Bengali and other Indian-language apps.
5. Possible benchmark leakage. Google warns that public benchmarks may have leaked into training and asks developers to validate on private data.
6. Small models struggle as agents. The paper says the 4B models "were poorly suited" to the AgentClinic agent benchmark (Nature Medicine, 6 Oct 2026).

What the licence actually says

The HAI-DEF Terms of Use (Google, last modified 15 Nov 2024) are the rules you accept when you download MedGemma. Key points in plain English:

1. You must not use it for anything that could make a health regulator treat Google as the "manufacturer" of a medical device, or in breach of law, or for uses on Google's Prohibited Use Policy.
2. If you build a product for diagnosis or treatment ("Clinical Use"), you must seek regulatory authorisation where applicable. In India that would mean checking with the CDSCO medical-device rules (our note, not legal advice).
3. If you share the model or offer it as a hosted service, you must pass on the use restrictions and a copy of the terms.
4. Google does not claim ownership of your outputs, but you are "solely responsible" for them, and the models come "as is" with no warranty.
5. The sample code is Apache 2.0; the model weights are not.

Why "run it yourself" matters for privacy in India

With a cloud AI API, you send each X-ray or patient note to someone else's server. With an open-weight model like MedGemma, you can run it on your own hardware, inside your hospital network, even fully offline. Google Research says MedGemma "runs locally on-device without internet access" (Google Research blog, 23 Oct 2026), and the Nature Medicine paper lists "the ability to run locally or offline" and "data and model sovereignty" as reasons to choose it over large API models (6 Oct 2026).

This is the same idea behind SR Infobiz's own free browser tools: the safest data is data that never leaves your control.

The DPDP Act angle

India's Digital Personal Data Protection (DPDP) Rules, 2025 were notified on 13 November 2025. The main duties for organisations, such as clear privacy notices, consent, security safeguards and breach reporting, take effect 18 months later, on 13 May 2027 (MeitY Gazette notification G.S.R. 846(E), 13 Nov 2025; LexCounsel, Nov 2025). KPMG notes that for healthcare this means explicit consent, purpose limitation and breach reporting (KPMG India, Dec 2025).

Self-hosting MedGemma does not make you DPDP-compliant by itself. You still need consent, security and governance. But it removes one big risk: shipping patient data to an outside AI vendor and its servers. That is our reading, not legal advice; talk to a privacy lawyer for your case.

Who in India is already using it?

Google's MedGemma lead author named AIIMS New Delhi among partners building on Google's HAI-DEF models (Andrew Sellergren, LinkedIn, 6 Oct 2026). Google Research also describes using MedSigLIP and MedGemma for tuberculosis screening tools that local teams can tune on modest regional datasets (Google Research blog, 23 Oct 2026). We could not confirm the details of any specific Indian deployment, so we are not naming products.

Hardware and cost to run

Google does not publish an official price list for self-hosting, so here is what is documented, plus our own rough estimate (clearly labelled):

1. The models are free to download from Hugging Face after you accept the HAI-DEF terms (Hugging Face, google/medgemma-4b-it).
2. Google's quick-start code runs the 4B model "locally on GPU" with the Hugging Face Transformers library (version 4.50.0 or newer) (Google, MedGemma model card).
3. Google notes you need Colab Enterprise to get enough GPU to run either 27B model without quantization (Hugging Face model card).
4. For scale, Google recommends deploying through Google Cloud Model Garden, where normal cloud charges apply (Google, model card).
5. The paper reports a 500-fold difference in compute cost between MedGemma 4B and the most expensive model it was compared with (Nature Medicine, 6 Oct 2026).

Our rough estimate (not tested): model weights in 16-bit format take about 2 bytes per parameter, so the 4B model needs roughly 8 GB of GPU memory just for weights and the 27B model roughly 54 GB, plus extra for images and long inputs. In practice that suggests a 16–24 GB GPU for the 4B model and a data-centre GPU (or a quantized version) for 27B. Quantized community versions exist on Hugging Face, but they are not Google's and their accuracy is not covered by Google's benchmarks.

How to try MedGemma (for developers)

1. Read the HAI-DEF Terms of Use and the model card first.
2. Go to the MedGemma collection on Hugging Face, sign in, and accept the access conditions.
3. Start with google/medgemma-1.5-4b-it in Google's quick-start Colab notebook.
4. Install Transformers (pip install -U transformers) and use the image-text-to-text pipeline from the model card.
5. Test only with public or synthetic data (for example, the CC0 chest X-ray in Google's sample code), never real patient data on a shared notebook.
6. If you plan a product, budget for fine-tuning on your own data, clinical validation and regulatory review.

If you only need image search or classification (no written answers), Google recommends MedSigLIP instead (Google, model card).

FAQ

Is MedGemma free?

The weights are free to download and use under the HAI-DEF Terms of Use. You pay for your own hardware or cloud compute (Google, HAI-DEF terms; model card).

Is MedGemma open source?

It is open-weight, not classic open source. The code is Apache 2.0, but the model itself has its own Google terms with use restrictions (Google, HAI-DEF terms).

Can a doctor or patient use MedGemma for diagnosis?

No. Google says its outputs are not meant to directly inform diagnosis or treatment and need independent verification (Google, model card). Always consult a qualified doctor.

Does MedGemma work offline?

Yes. Once downloaded, it can run on your own machine without internet (Google Research blog, 23 Oct 2026).

Does it understand Hindi or other Indian languages?

Google's safety testing was "primarily English", and we found no official Indian-language benchmarks. Test carefully before using it in regional languages.

Which version should I start with?

MedGemma 1.5 4B for most experiments; 27B if you need the strongest text reasoning and have the hardware (Google, model card).

Disclaimer: This article is for general information about AI technology. It is not medical, legal or regulatory advice. SR Infobiz has not tested MedGemma; all performance figures are as reported by the named sources.

Sources

All accessed 24 October 2026.

1. Sellergren, A., Kazemzadeh, S. et al., "An open vision-language model for diverse medical applications", Nature Medicine, 6 Oct 2026 — https://www.nature.com/articles/s41591-026-04626-w
2. Google for Developers, "MedGemma 1.5 model card" — https://developers.google.com/health-ai-developer-foundations/medgemma/model-card
3. Google for Developers, "MedGemma" overview — https://developers.google.com/health-ai-developer-foundations/medgemma
4. Hugging Face, google/medgemma-4b-it model card — https://huggingface.co/google/medgemma-4b-it
5. Google, "Health AI Developer Foundations Terms of Use", last modified 15 Nov 2024 — https://developers.google.com/health-ai-developer-foundations/terms
6. Google Research blog, "Bringing foundational models to the front lines of TB screening", 23 Oct 2026 — https://research.google/blog/bringing-foundational-models-to-the-front-lines-of-tb-screening/
7. AI Weekly, "Google releases MedGemma, an open medical vision-language model line", 20 Oct 2026 — https://aiweekly.co/alerts/google-releases-medgemma-an-open-medical-vision-language-model-line
8. Andrew Sellergren, LinkedIn post, 6 Oct 2026 — https://www.linkedin.com/posts/asellerg_an-open-vision-language-model-for-diverse-activity-7513272294144675842-BsuD
9. MeitY, Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), 13 Nov 2025 — https://www.dpdpa.in/dpdpa_rules_2025/dpdpa_rules_english_2.pdf
10. LexCounsel, "Regulatory Update: Notification of the DPDP Rules 2025", Nov 2025 — https://lexcounsel.in/newsletters/regulatory-update-notification-of-the-digital-personal-data-protection-rules-2025/
11. KPMG India, "The privacy prescription: Impact of DPDP Act and rules in healthcare and life sciences sector", Dec 2025 — https://assets.kpmg.com/content/dam/kpmgsites/in/pdf/2025/12/the-privacy-prescription-impact-of-dpdp-act-and-rules-in-healthcare-and-life-sciences-sector.pdf.coredownload.pdf