Google releases MedGemma, open medical vision-language models

Overview

Updated Oct 20

Google's MedGemma family, published in Nature Medicine this month, scored 86.2% on MedQA, a standard medical question benchmark. The open models also beat their Gemma 3 base by 2.6% to 10% on medical image questions and 15.5% to 18.1% on chest X-ray finding classification.

Google says the models have been downloaded more than 10 million times by over 65,000 developers. Clinics are using the open weights in places an API would not reach: Crane AI runs an adapted model on-device in rural Uganda for triage without internet, and Indonesia's health ministry is building a tuberculosis detector from local chest X-rays, aiming to screen 50 million people a year.

Why it matters

Hospitals can run state-of-the-art medical AI on their own servers, keeping patient data in-house instead of sending it to a closed API.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

86.2%

MedQA score

Best open-model result at release on the MedQA medical question benchmark.

81%

Clinically equivalent or superior reports

Share of MedGemma-generated chest X-ray reports that clinicians rated likely to support the same or better clinical decisions.

Model variants

4B and 27B multimodal models, a 27B text model, and the MedSigLIP vision encoder.

400M

MedSigLIP parameters

Size of the medically tuned vision encoder derived from SigLIP.

10M+

Downloads since release

More than 10 million downloads by over 65,000 developers, reported by Google at its Research@ 2026 event on October 14.

50M

Indonesia TB screening target

Annual screening goal of Indonesia's health ministry, which is building a MedGemma and MedSigLIP tuberculosis detector on local chest X-ray data.

Voices

Curated perspectives — historical figures and your fellow readers.

Organizations Involved

Google Research

Corporate research division

Lead developer of MedGemma

Google's research arm, which co-developed MedGemma with Google DeepMind.

Google DeepMind

AI company

Co-developer of MedGemma

Google's AI lab, which co-developed MedGemma and contributed the Gemma 3 base models.

Nature Medicine

Peer-reviewed journal

Publisher of the MedGemma paper

A leading medical research journal that published the MedGemma paper on October 6, 2026.

Timeline

4 events Latest: October 16th, 2026 · 4 weeks ago

Tap a bar to jump to that date

October 2026
Google publicly releases MedGemma
Latest Release

Google announces the open model family: 4B and 27B multimodal variants, a 27B text model, and the MedSigLIP encoder.
October 16th, 2026
Google announces $400,000 MedGemma grant in India
Funding

Google announced a $400,000 grant to support MedGemma adoption and research in India, part of its AI initiatives in the country.
October 16th, 2026
Google reports 10 million MedGemma downloads
Adoption

At its Research@ 2026 event, Google said MedGemma had been downloaded more than 10 million times by over 65,000 developers since its Nature Medicine publication. Researchers are working to connect the models with electronic health records and to apply them to mammography, ultrasound, and electrocardiogram data.
October 14th, 2026
Nature Medicine publishes MedGemma paper
Publication

Peer-reviewed paper describes the model family, benchmark results, and the MedSigLIP vision encoder.
October 6th, 2026

Scenarios

Hospitals deploy MedGemma on-premises for radiology

Likely Resolves by End of 2027

Discussed by: AI Weekly and healthcare technology analysts

Hospitals handling protected health information that can't cross API boundaries adopt MedGemma for chest X-ray and pathology workflows. The open weights allow on-prem deployment and fine-tuning on local data. The paper's 81% clinically equivalent or superior report rate gives radiology departments a concrete quality benchmark to justify adoption.

MedGemma-based tools enter FDA regulatory pathway

Possible Resolves by Q2 2028

Discussed by: Healthcare regulatory analysts

Companies build MedGemma-based diagnostic tools and submit them for FDA clearance. The open weights and published validation data could accelerate the regulatory process, since the model's performance is already documented in a peer-reviewed journal.

MedGemma becomes the base for a wave of fine-tuned medical models

Likely Resolves by Q2 2027

Discussed by: The Nature Medicine paper authors

The paper's central claim is data efficiency: fine-tuning MedGemma beats fine-tuning base Gemma 3 for medical tasks, especially with limited training data. If that holds, MedGemma becomes the default starting point for specialized medical models, from pathology classifiers to EHR information retrieval systems.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

May 2023

Med-PaLM 2 (2023)

Google unveiled Med-PaLM 2, a medical AI model that scored 86.5% on MedQA, but kept it behind a closed API for a limited test group of clinicians.

Then

Clinicians could test the model but not deploy it on their own infrastructure or fine-tune it on local data.

Now

Set the pattern of closed medical AI that MedGemma now breaks.

Why this matters now

MedGemma is Google's open answer to its own closed Med-PaLM 2, achieving a comparable MedQA score while being freely downloadable.

November 2023

LLaVA-Med (2023)

Researchers released LLaVA-Med, an early open vision-language model for medical imaging, built by fine-tuning the LLaVA model on biomedical image-text data.

Then

Demonstrated that open medical VLMs were feasible, though accuracy lagged far behind closed models.

Now

Established the open medical VLM research line that MedGemma extends with production-grade performance.

Why this matters now

Shows the progression from research prototype to production-grade open medical model in roughly three years.