Multi-Modal Foundation · google
MedGemma 27B
An instruction-tuned 27B medical AI for clinical question answering, health records, and medical image interpretation.
What MedGemma 27B is
MedGemma 27B (google/medgemma-27b-it) is an open multimodal generative artificial intelligence model developed by Google. Built on the Gemma 3 architecture, it is tailored for medical text, electronic health records, and medical image comprehension. The model serves as an adaptable foundation to help developers and researchers build healthcare-oriented applications.
The model addresses the challenge of interpreting complex medical data across distinct modalities, including clinical literature, patient records, and biomedical imaging. By combining text generation with medical visual encoding, it provides baseline capabilities for tasks spanning multiple clinical specialities, including radiology, histopathology, ophthalmology, and dermatology.
MedGemma 27B multimodal incorporates pre-training across medical images, clinical records, and medical comprehension tasks, including Fast Healthcare Interoperability Resources (FHIR) data. It is instruction-tuned and intended as a starting point for downstream adaptation, validation, and domain-specific fine-tuning rather than unverified direct clinical deployment.
How it works
The multimodal pipeline pairs this language model with a specialised SigLIP image encoder that processes input images normalised to 896 × 896 pixels, encoding each image into 256 tokens. The model generates text outputs of up to 8,192 tokens.
The model's training was executed using JAX on TPUs, starting from base Gemma 3 weights pre-trained on text and code. The SigLIP vision encoder was pre-trained on de-identified radiology, histopathology, dermatology, and ophthalmology data. The unified model was further trained on diverse medical datasets, including clinical question-answer sets, FHIR-based synthetic health records, radiology datasets with bounding boxes such as Chest ImaGenome, and proprietary histopathology whole-slide images and clinical photography.
Key facts
- Model type
- Decoder-only Transformer (Gemma 3)
- Attention mechanism
- Grouped-query attention (GQA)
- Base model
- google/gemma-3-27b-pt
- Output length
- Up to 8,192 tokens
- Input modalities
- Text and vision (images normalised to 896 × 896 resolution)
- Output modality
- Text
- Image encoder
- SigLIP (pre-trained on medical imaging data)
- Model version
- 1.0.0
- Creation date
- 9 July 2025
What you can use it for
- Medical visual question answering
- Answering questions grounded in single medical images, including radiographs and clinical photographs, evaluated across benchmarks such as SLAKE and VQA-RAD.
- Electronic health record comprehension
- Interpreting and answering multi-faceted queries regarding patient clinical history using structured, FHIR-based electronic health record data.
- Medical image interpretation and classification
- Categorising clinical images across modalities, including chest X-rays (MIMIC-CXR, CheXpert), skin lesions (PAD-UFES-20, SCIN), fundus photography (EyePACS), and histopathology slides (CAMELYON, TCGA).
- Clinical question answering and reasoning
- Responding to complex medical and biomedical examinations and clinical queries across benchmarks such as MedQA, MedMCQA, PubMedQA, and MedXpertQA.
- Chest X-ray report generation
- Drafting descriptive findings and reports from chest radiographs, which can be further adapted and fine-tuned to match specific institutional reporting formats.
Run it here
We list this model but cannot run it yet: gated on the Hub — never run past a gate the user did not accept.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Exhibits specialised comprehension of structured FHIR electronic health record data, scoring 90.5 on EHRQA benchmarks.
- Incorporates a SigLIP vision encoder pre-trained on diverse medical imagery, including dermatology, fundus photography, radiology, and whole-slide histopathology.
- Demonstrates strong text-based medical knowledge, achieving 87.0% on MedQA (best-of-5) and 86.2% on MMLU Med.
Watch out for
- Not intended to directly guide clinical diagnosis, patient management, or treatment decisions without independent validation and human clinician correlation.
- Multimodal capabilities are primarily evaluated on single images and have not been evaluated for multi-image comprehension tasks.
- The model has not been evaluated or optimized for multi-turn conversational interactions.
- Model performance can be more sensitive to specific prompt formulations compared to the base Gemma 3 model.
- Instruction tuning alters reporting style compared to standard ground truth datasets, resulting in lower out-of-the-box RadGraph F1 scores for CXR report generation unless specifically fine-tuned.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use MedGemma 27B commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Can this model directly assist in clinical patient care?
No. The outputs are not intended to directly guide clinical diagnosis, patient management, or treatment decisions. All generated outputs are preliminary and require independent verification and clinical correlation.
How are images processed by MedGemma 27B?
Images are normalised to an 896 × 896 resolution and encoded into 256 tokens using a medically pre-trained SigLIP image encoder before being passed into the language model.
Can MedGemma 27B analyse multiple images at once?
No. The model's multimodal performance has been primarily evaluated on single-image tasks and has not been tested or evaluated on inputs containing multiple images.
What is the difference between MedGemma 27B multimodal and the 27B text-only version?
The multimodal version accepts both vision and text inputs and has specific pre-training on FHIR records and medical images. The text-only version accepts only text, performs slightly higher on certain text-only medical benchmarks, and lacks image comprehension capabilities.
Is the model optimized for multi-turn dialogue?
No. The model has not been evaluated or optimized for multi-turn conversational applications.
How to cite it
Sellergren et al. "MedGemma Technical Report." arXiv preprint arXiv:2507.05201 (2025).
Written from MedGemma 27B's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.