What is MedGemma 1.5 4B?
MedGemma 1.5 4B is Google's open medical vision-language model, released on 13 January 2026 as an update to the MedGemma 4B of May 2025. It is built on Gemma 3 with a SigLIP image encoder pre-trained on de-identified medical images, and it answers in text. It is offered only as an instruction-tuned 4B model.
The update adds inputs the first version could not handle: 3D CT and MRI volumes, several patches from one whole-slide pathology image, chest X-rays compared against earlier scans, bounding boxes around anatomy, and structured values pulled from lab reports and health records. CT, MRI and slide images need pre-processing first. Each image is resized to 896 by 896 pixels and becomes 256 tokens, and the context holds up to 128K tokens.
Use is governed by Google's Health AI Developer Foundations (HAI-DEF) terms, not Apache 2.0, and the Hugging Face download is gated behind accepting them. In the last 30 days it drew about 198,000 downloads.
Is MedGemma 1.5 4B worth it?
MedGemma 1.5's reason to exist is breadth of medical input at 4B parameters. Google calls it the first open medical generalist model to read full CT and MRI volumes natively. On Google's card, MRI condition accuracy rises from 51.3 to 64.7 and CT from 58.2 to 61.1 over MedGemma 1, whole-slide pathology ROUGE jumps from 2.2 to 49.4, and the EHRQA records test goes from 67.6 to 89.6. Chest ImaGenome anatomy localisation rises from 3.1 to 38.0 IoU.
The gains are uneven. SLAKE visual Q&A falls from 72.3 to 59.7, which Google puts down to the new model being less tuned for that format, PubMedQA drops from 73.4 to 68.2, and chest X-ray report generation scores 27.2 against 30.3 for a MedGemma 1 version fine-tuned for that job. Absolute scores on hard tests stay low: 20.9 on multimodal MedXpertQA and 27.0 macro F1 on CT-RATE.
Google is explicit about the limits. The card says outputs are not intended to directly inform diagnosis, treatment or patient management without validation and adaptation, and that the model was mostly evaluated on single images, not multi-image or multi-turn use. The HAI-DEF terms put regulatory approval on you, require that you not use it in a way that would make Google a medical-device manufacturer, and require passing the use restrictions to anyone you redistribute to.
As a base it is strong value: 4B parameters fit on one modest GPU, so hospital IT teams can run and fine-tune it on premises where patient data cannot leave.
Where it earns its place
A clinical informatics team at a regional hospital
Pull lab values and units out of scanned outside reports into the records system
Run the model on premises and prompt for structured output; Google reports macro F1 rising from 78 to 91 on one EHR extraction set. A clinician still reviews each import, as the terms and card expect.
A machine-learning engineer at a radiology AI start-up
Build a triage prototype that flags CT studies for review
Fine-tune on your own labelled volumes after Google's pre-processing step. The base 61.1 CT accuracy is a starting point, not a product; regulatory clearance is your responsibility under HAI-DEF.
A digital pathology researcher at a university hospital
Draft descriptions of whole-slide images for a study dataset
Feed several patches per slide; WSI ROUGE rose from 2.2 to 49.4 in Google's table. Use outputs as drafts for pathologists, since the card warns against direct clinical use.
Before you commit
- Setup
- Accept the HAI-DEF terms on Hugging Face, then load with transformers; or deploy from Vertex AI Model Garden. CT, MRI and whole-slide inputs must be pre-processed first, as the card describes.
- Integrations
- Hugging Face transformers, Vertex AI, MedSigLIP for image-only tasks, and Google's MedASR speech model for dictated prompts.
- Data and privacy
- Self-hosting keeps patient data on your hardware. HAI-DEF terms make you responsible for regulatory approval and require deleting copies if the agreement ends.
- Cost at scale
- Weights are free; a 4B model runs on a single GPU. Vertex AI deployment is billed at standard Google Cloud compute prices.
- Limits
- Output up to 8,192 tokens; mainly evaluated on single images and single turns; more sensitive to prompt wording than Gemma 3, per the card; safety testing was mostly in English.
What’s good
- Reads 3D CT and MRI volumes and whole-slide pathology
- MRI accuracy up from 51.3 to 64.7 over MedGemma 1
- EHRQA up from 67.6 to 89.6; strong on lab-report extraction
- 4B size runs on premises, so patient data can stay local
What’s not
- Not cleared for clinical decisions without your own validation
- SLAKE and PubMedQA scores fell from version 1
- HAI-DEF terms, not Apache 2.0, with flow-down obligations
- CT, MRI and slide inputs need pre-processing
Best for
Health-tech and hospital engineering teams building and validating their own imaging or records tools on premises.
MedGemma 1.5 4B alternatives
MedGemma 1.5 4B: common questions
- Is MedGemma 1.5 4B worth it?
- MedGemma 1.5 is the open starting point for medical imaging apps that need 3D scans, but it is a base to validate and adapt, not a model to put in front of patients. We rate it 3.9 out of 5.
- How much does MedGemma 1.5 4B cost?
- Free weights under the Health AI Developer Foundations terms. Vertex AI hosting is billed at Google Cloud compute rates.
- What are the downsides of MedGemma 1.5 4B?
- Not cleared for clinical decisions without your own validation; sLAKE and PubMedQA scores fell from version 1; hAI-DEF terms, not Apache 2.0, with flow-down obligations; cT, MRI and slide inputs need pre-processing.
- Who is MedGemma 1.5 4B best for?
- Health-tech and hospital engineering teams building and validating their own imaging or records tools on premises.
- How much does MedGemma 1.5 4B cost for a team or in production?
- Weights are free; a 4B model runs on a single GPU. Vertex AI deployment is billed at standard Google Cloud compute prices.
- Is MedGemma 1.5 4B safe for company data?
- Self-hosting keeps patient data on your hardware. HAI-DEF terms make you responsible for regulatory approval and require deleting copies if the agreement ends.
- What are the best MedGemma 1.5 4B alternatives?
- The closest alternatives we have reviewed are Gemma 4, Qwen3.8-27B, MiniCPM-V 4.7 (35B-A3B and 1B) and Phi-4-multimodal.
Answers checked Nov 7, 2026.