MedGemma 4B
Google's open multimodal medical model: 4B parameters on Gemma 3 for radiology, clinical text and EHR tasks, under HAI-DEF terms.
About
MedGemma 4B is Google's compact open medical model, launched at Google I/O in May 2025 as part of the Health AI Developer Foundations (HAI-DEF) program. It is a multimodal 4B-parameter model built on Gemma 3, designed for tasks like chest X-ray interpretation, radiology report drafting, clinical text summarization and structured data extraction. Google reported 64.4% on MedQA for the 4B variant and, notably, that a US board-certified radiologist judged its generated chest X-ray reports sufficient for patient management in 81% of cases.
With over 660,000 downloads it is by far the most-adopted open medical model, and it runs on modest hardware including single GPUs and mobile-class devices via Ollama. Google followed it with a 27B text model and MedGemma 1.5 in January 2026, which added 3D CT/MRI and whole-slide pathology support. Two things to keep straight: the license is Google's HAI-DEF terms, not Apache 2.0, and Google is explicit that MedGemma is a developer building block, not a validated medical device. Nothing here should be read as clinical validation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Google-reported: MedGemma 4B reaches 64.4% on MedQA (provider claim)
- Google-reported: chest X-ray reports judged sufficient for patient management in 81% of cases by a US board-certified radiologist (provider claim)
- Press-reported 88.9 F1 on chest X-ray analysis (PakTribune, citing Google)
- Google stresses MedGemma is not a validated medical device; independent clinical evals not verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: launched at Google I/O May 2025 (press coverage dated May 23, 2025); HF repo google/medgemma-4b-it created May 19, 2025 with 667,706 downloads and 1,101 likes (HF API, Oct 2026); Google-published figures: 64.4% MedQA, 81% radiologist-sufficient CXR reports (via BiopharmaTrend/MarkTechPost coverage of Google's materials). Open weights. Strengths: dominant adoption in open medical AI, real vendor investment (1.5 release Jan 2026), runs on modest hardware. Docked for: (1) all clinical-leaning figures are Google's own; (2) HAI-DEF terms are custom, not an open license, and gated behind terms acceptance; (3) Google explicitly disclaims clinical validity. License: HAI-DEF (Health AI Developer Foundations terms, custom).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Launched at Google I/O May 2025; HF repo google/medgemma-4b-it created 2025-05-19; 667,706 downloads, 1,101 likes (HF API Oct 2026)
- · Google-reported: 64.4% MedQA; CXR reports sufficient for patient management in 81% of cases (radiologist-judged)
- · Followed by 27B model and MedGemma 1.5 (Jan 2026)
- · HAI-DEF terms custom and gated; Google disclaims clinical validity