MedGemma 27B (multimodal)
Release in the MedGemma family · version 1.0.0 (27B multimodal, instruction-tuned)
Maintained by Google (Health AI Developer Foundations)1
MedGemma 27B multimodal is the 27B multimodal variant of MedGemma 1, released by Google on July 9, 2025. It accepts text and images and produces text, and unlike the 27B text-only variant it was also trained on medical images and FHIR-based electronic health record data. It is published only as an instruction-tuned model.18
- Documentation: MedGemma 1 model card (Health AI Developer Foundations) (external site: developers.google.com)
- Model hub: Hugging Face repository (external site: huggingface.co)
- Paper: MedGemma Technical Report (arXiv 2507.05201) (external site: arxiv.org)
- License: Health AI Developer Foundations terms of use (external site: developers.google.com)
Availability and license
Overall availability
Downloadable from Hugging Face after logging in and acknowledging the Health AI Developer Foundations terms of use; the gate states that requests are processed immediately. Also offered through Google Cloud Model Garden.321
Availability is separate from permission: read the license before using or redistributing.
Health AI Developer Foundations Terms of Use (external site: developers.google.com)314
Apache License 2.0 (google-health/medgemma notebooks and serving code) (external site: raw.githubusercontent.com)56
The HAI-DEF terms (licensor Google LLC) permit use, modification, and distribution subject to use restrictions, including the HAI-DEF Prohibited Use Policy and a bar on uses that could lead a health regulator to deem Google a medical device manufacturer. Redistributors must pass on the terms and use restrictions and include a specified notice file. The terms require indemnification of Google and are governed by California law.4
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Weights are on Hugging Face behind a click-through acknowledgment of the HAI-DEF terms that is processed automatically.32 |
| Inference codeIs code for running the model published? | Public | The card documents inference with Hugging Face Transformers (4.50.0 or later), and the Hugging Face page also shows vLLM and SGLang serving; notebooks and serving code are in the Apache-2.0 google-health/medgemma repository.125 |
| Training codeIs the code used to train the model published? | Unknown | The card states training was done with JAX; fine-tuning notebooks are published, but the code used to train the model was not found.15 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The card lists public datasets (for example MIMIC-CXR and SLAKE) and describes licensed or internally collected de-identified datasets that are not public.1 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The MedGemma Technical Report describes the models and their training at a summary level.7 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The card publishes results on public and internal benchmarks; for 27B results it notes test-time scaling was used. Several evaluation datasets are internal.1 |
What it is useful for
Run and use notes
- The card documents inputs of up to 128K tokens with images normalized to 896 x 896 and encoded to 256 tokens each, and outputs of up to 8,192 tokens.1
Organization context
Provenance and derivatives
Built by Google on Gemma 3 27B; Hugging Face metadata lists google/gemma-3-27b-pt as the base model. The multimodal variant uses a SigLIP image encoder pre-trained on de-identified medical data.31
- Derived from: Gemma 3 27B (pre-trained), google/gemma-3-27b-pt — Base model; see the Gemma family record.
Other releases in the MedGemma family
- MedGemma 1.5 4BModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: Documented U.S. control
The model card lists Google as author, and the governing terms are issued by Google LLC. Alphabet Inc.'s fiscal 2025 Form 10-K lists its principal executive offices in Mountain View, California, and Exhibit 21.01 lists Google LLC as a Delaware subsidiary of Alphabet.14910
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.