Multi-Modal Foundation · google
MedGemma 1.5 4B
A 4-billion-parameter multimodal model interpreting medical imaging and clinical text for biomedical researchers.
What MedGemma 1.5 4B is
MedGemma 1.5 4B is an instruction-tuned multimodal model developed by Google to interpret medical text and biomedical images. Built upon the Gemma 3 architecture, the model serves as an open foundation designed to assist researchers and software developers in building healthcare AI applications.
The model expands upon earlier releases by adding capability for high-dimensional clinical inputs. These include three-dimensional computed tomography (CT) and magnetic resonance imaging (MRI) volumes, simultaneous multi-patch inputs from whole-slide histopathology images, longitudinal chest X-ray comparisons against historical scans, and bounding box localization of anatomical features.
In addition to multimodal imaging tasks, MedGemma 1.5 4B interprets text-based electronic health record (EHR) data and converts unstructured medical documents, such as laboratory reports in PDF or image format, into structured JSON tables.
How it works
MedGemma 1.5 4B uses a decoder-only transformer architecture equipped with grouped-query attention (GQA), adapted from Gemma 3. For visual comprehension, it incorporates a SigLIP vision encoder pre-trained on de-identified clinical data spanning radiographs, histopathology, dermatology, and retinal fundus photography.
The model was trained using JAX on tensor processing units (TPUs). Its language component was trained on public and licensed medical datasets, including clinical question-answering pairs, biomedical literature from PMC-OA, de-identified electronic health records, 2D and 3D radiological scans, multi-stain whole-slide histopathology, and laboratory reports.
Key facts
- Architecture
- Decoder-only Transformer with Grouped-Query Attention (GQA) and SigLIP vision encoder
- Parameter count
- 4 billion (4B)
- Input modalities
- Text, vision
- Output modality
- Text only
- Image encoding resolution
- 896 × 896 pixels (256 tokens per image)
- Release date
- 13 January 2026
What you can use it for
- 3D Volumetric Scan Interpretation
- Interpreting three-dimensional volume representations of axial CT and MRI studies across anatomical regions such as the head, chest, and abdomen to classify conditions.
- Whole-Slide Histopathology Analysis
- Processing multiple image patches from whole-slide histopathology scans concurrently to evaluate diagnostic text and cancer grading.
- Longitudinal Radiograph Tracking
- Comparing current chest radiographs directly against historical prior scans to assess disease progression.
- Anatomical Feature Localization
- Detecting anatomical landmarks and pathological findings on chest X-rays by generating bounding box coordinate descriptions.
- Clinical Document Data Extraction
- Extracting structured clinical variables, values, and measurement units from unstructured PDF or image laboratory reports into JSON format.
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
- Supports 3D CT and MRI volumetric inputs alongside multi-patch digital histopathology and standard 2D clinical photography.
- Utilises a domain-specific SigLIP image encoder pre-trained on diverse de-identified medical and pathological imaging datasets.
- Delivers improved accuracy on medical question answering and structured document extraction over base Gemma 3 and predecessor models.
Watch out for
- Outputs must not directly inform clinical diagnoses, patient management decisions, or treatment recommendations without downstream adaptation and verification.
- The model has not been evaluated or optimised for multi-turn conversational interactions.
- Displays greater sensitivity to prompt phrasing compared to the base Gemma 3 architecture.
- Evaluations were conducted primarily on single-image prompts and English-language benchmarks.
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 1.5 4B 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 MedGemma 1.5 4B analyse 3D radiological images?
Yes. It supports volumetric representations of axial CT and MRI studies, though these inputs require pre-processing prior to ingestion.
Does the model generate segmented images or visual overlays?
No. The model has a text-only output modality; anatomical localization tasks are represented as textual bounding box coordinates.
How are visual inputs processed?
Images are normalised to an 896 × 896 pixel resolution and mapped by the SigLIP vision encoder into 256 tokens per image.
Can the model extract data from non-digital lab records?
Yes. MedGemma 1.5 4B can process unstructured laboratory report documents in PDF or PNG formats and output the extracted test data into structured JSON format.
How to cite it
Sellergren et al. "MedGemma 1.5 Technical Report." arXiv preprint arXiv:2604.05081 (2026).
Written from MedGemma 1.5 4B'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.