News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgMedGemma

MedGemma

Medical vision-language models built on Gemma 3 for images and text

MedGemma is a collection of medical vision-language foundation models built on Gemma 3 4B and 27B, announced in July 2025. Google researchers report it handles medical images and text with strong reasoning while keeping the general abilities of the Gemma 3 base models, and it is released with weights and tutorials.

Share on WhatsApp

  • Built on Gemma 3, in 4B and 27B versions
  • Chest X-ray finding classification improved 15.5-18.1% over base models
  • Fine-tuning cut health record retrieval errors by 50%
  • MedSigLIP, a medical vision encoder derived from SigLIP, powers its image understanding
  • Technical report first posted 7 July 2025, revised April 2026

What it is

MedGemma is a set of medical foundation models that work on both images and text. It comes in versions derived from Gemma 3 at 4B and 27B parameters. The team behind it is a large group of Google researchers, who published a technical report in July 2025. The models are aimed at healthcare tasks such as reading scans, answering medical questions and pulling information from health records. The developers say MedGemma beats similar-sized generative models on medical work and comes close to models built for a single task, while still holding on to the general abilities of the Gemma 3 models it is based on.

How it came about

The MedGemma technical report was submitted on 7 July 2025 and later revised, with the fourth version dated 6 April 2026. It carries more than eighty author names, led by Andrew Sellergren, and covers the whole collection rather than one model. Alongside the language models, the team introduced MedSigLIP, a vision encoder adapted from SigLIP for medical use. Model weights and tutorials were made available together with the report.

Inside the model

MedGemma is built on Gemma 3, in 4B and 27B sizes, and is tuned for medical images and text. Its visual side runs through MedSigLIP, a vision encoder derived from SigLIP and adapted to medical imagery. The report says the models keep the general capabilities of their Gemma 3 bases. Fine-tuning on a subdomain brings further gains: the team reports cutting errors in health record information retrieval by half, and reaching performance on a par with specialised methods for pneumothorax classification and histopathology patch classification.

What it can do

Against the Gemma 3 base models on tasks outside the training distribution, the team reports gains of 2.6 to 10 per cent on medical multimodal question answering, 15.5 to 18.1 per cent on classifying findings in chest X-rays, and 10.8 per cent on agentic evaluations. MedSigLIP, used on its own as an encoder, is said to match or beat specialised medical image encoders. The report frames these as improvements over the base models and comparisons with task-specific systems, not as universal rankings.

How to use it

The MedGemma collection is published with model weights and tutorials. The technical report points readers to an online location for both.

Why it matters

The developers present MedGemma as a base for building healthcare AI, arguing that a model already good at medical tasks needs less task-specific tuning data. That, they say, could speed up medical research and the development of downstream applications, given how varied healthcare data is and how much privacy matters in the field.

Questions people ask about MedGemma

What is MedGemma?

A collection of medical vision-language foundation models built on Gemma 3 at 4B and 27B parameters. Google researchers released it in July 2025 for work with medical images and text, such as answering questions about scans and reading health records.

How does it differ from Gemma 3?

It starts from Gemma 3 and is tuned for medicine. The developers say it keeps the general abilities of the base models while improving on medical tasks, and its image understanding runs through MedSigLIP, a vision encoder adapted from SigLIP for medical imagery.

How well does it perform?

The technical report gives gains over the base models of 2.6 to 10 per cent on medical multimodal question answering, 15.5 to 18.1 per cent on chest X-ray finding classification, and 10.8 per cent on agentic evaluations. These are the team's own reported figures.

Can I use it?

Yes. Model weights and tutorials are published with the technical report. The report directs readers to an online location for the collection.

What is MedSigLIP?

A vision encoder derived from SigLIP and tuned on medical data. It supplies the visual understanding inside MedGemma, and the developers say that used alone it matches or beats specialised medical image encoders.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MedGemma yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2507.05201. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.