AI for Bio · Model

MedGemma / Med-Gemini

Google medical language-model family: closed Med-Gemini and open-weight MedGemma

Text, LLMs & agentsBig-tech labUnited StatesActive中文

Overview

NameMedGemma / Med-Gemini · aliases: MedGemma, Med-Gemini, MedGemma 1.5
Full nameNo separate expanded name
Ambiguous names (not auto-linked)Gemma
FamilyCovers closed Med-Gemini (Gemini-based) and open-weight MedGemma versions (Gemma 3-based); dates and openness differ
  • Med-Gemini · †
  • MedGemma 4B / 27B · †
  • MedGemma 1.5 4B ·
OrganizationGoogle Research, Google DeepMind
Organization typeBig-tech lab
CountryUnited States
Model classMedical multimodal LLM
DomainText, LLMs & agents
First release† partly verified · preprint
Basis: arXiv v1 of the earliest member, Med-Gemini
Note: This date is Med-Gemini (no public weights) while the latest update is MedGemma; read per member source
Other milestones
  • Preprint v1: † · Med-Gemini arXiv v1 source
  • Journal online: · MedGemma published in Nature Medicine source
  • Weights release: · MedGemma 1.5 4B weights source
Latest update · journal
Basis: MedGemma published in Nature Medicine
MedGemma paper in Nature Medicine; MedGemma 1.5 4B came out 2026-01-13; no MedGemma 2 found source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · New weights and journal publication within 12 months
OpennessCode: Partly open
Weights: Open · Health AI Developer Foundations terms (not Apache-2.0)
MedGemma weights open under Google terms, not Apache-2.0; Med-Gemini closed
Benchmarks reported1 counted (self 1, unconfirmed 1)
Wet-lab validationNone recorded
Unverified fields† first_release, benchmarks_reported[1]
As of2026-11-16

What it does

Google medical language models: MedGemma (4B multimodal, 27B text) is built on Gemma 3 with open weights; the earlier Med-Gemini is built on Gemini and closed. This site uses them mainly as biomedical QA comparators.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
14 PubMedQADeveloper report, official protocol—Self-reportedMedGemma 4B, 27B and 1.5 reported; Med-Gemini does not report PubMedQA. · evidence
23 BioASQUnconfirmed (not counted)—Unconfirmed sourceThe overview scores whole systems and contains no standalone figure for it, so the board marking in the matrix is unsupported. †disputed · evidence

Other evaluations (outside the 44)

None recorded.

Developers report: Source states no developer claim

Independent evidence: No independent evidence found. No independent evidence found

Training data

Group in Figure 7: Private or licensed data.

ClassValueDetail
a. Public databasesYesPublic medical text and images, including MedQA, MedMCQA and PubMedQA training splits
b. Private or licensed dataYesLarger internal non-public data, e.g. ~32.5M pathology patch-text pairs; licensing not stated
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Google Research and Google DeepMind (merged 2023-04) are part of Alphabet; MedGemma belongs to the Health AI Developer Foundations program
  • Funding: Not disclosed or not applicable
  • Products and partnerships: No product or partnership found
  • Related models: TxGemma, C2S-Scale, Frontier general-purpose LLMs

Reading notes

  1. First release and latest update belong to different members; not one model's timeline
  2. The BioASQ pairing is unconfirmed; only PubMedQA counts

Caveats

  1. Family entry: Med-Gemini is closed with no weights; MedGemma has open weights
  2. MedGemma training includes PubMedQA training splits; read its numbers with that in mind

Sources

  1. Crossref record 10.1038/s41591-026-04626-w (journal)
  2. arXiv preprint 2507.05201 (preprint)
  3. Hugging Face page google/medgemma-1.5-4b-it (repository)
  4. Official page research.google (company blog)
  5. arXiv preprint 2404.18416 (preprint)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.