Google MedGemma

Google MedGemma is Google’s open model family for medical text and image analysis, built on Gemma 3 and part of the Health AI Developer Foundations (HAI-DEF). The current model is MedGemma 1.5 4B from 13 January 2026, a multimodal model that, compared with MedGemma 1, additionally handles CT and MRI volumes, whole-slide pathology and chest X-ray time series. For complex text tasks, MedGemma 27B from the first generation remains available; there is no 27B variant of MedGemma 1.5.

innFactory AI Consulting from Rosenheim, Germany advises enterprises across the DACH region (Germany, Austria, Switzerland) on GDPR-compliant integration of medical AI. For patient data we strongly recommend self-hosting and careful compliance review under MDR/IVDR and the EU AI Act. This overview reflects the state as of October 2026.

Important Notice

MedGemma is NOT an approved medical device. It must not be used as the sole basis for clinical decisions. All results must be validated by qualified medical professionals.

Current models at a glance

MedQA figures (4 answer options) per the model card, for MedGemma 27B (text) per the Google Research blog. The model card lists 128K input tokens and 8,192 output tokens. All variants are covered by the Health AI Developer Foundations Terms of Use, not Apache 2.0.

New Capabilities in MedGemma 1.5

Enhanced Imaging Support

• 3D medical imaging: Processing complete CT and MRI volumes, not just individual slices
• Whole-slide pathology: Analysis of large-scale histopathology images through patch-based processing
• Longitudinal image analysis: Comparison of image series over time, such as multiple chest X-rays of the same patient
• Improved anatomical localisation: Significantly more precise identification of anatomical structures, particularly in chest radiographs

Performance compared with MedGemma 1 4B

Per Google, compared with MedGemma 1 4B:

• CT findings: accuracy from 58% to 61%
• MRI findings: accuracy from 51% to 65%
• Lab reports: macro F1 for structured data extraction from 60% to 78% (internal benchmark)
• Histopathology: ROUGE-L from 0.02 to 0.49
• Radiograph localisation: intersection-over-union from 3% to 38%
• Health records: EHRQA from 68% to 90%
• Medical knowledge: MedQA from 64% to 69%

MedGemma 1.5 4B or MedGemma 27B?

• MedGemma 1.5 4B: for imaging (CT, MRI, pathology, X-ray), document extraction and health records; compute-efficient and able to run offline
• MedGemma 27B: per Google still the choice for more complex text-based applications, with 87.7% MedQA (text) or 85.3% (multimodal)

Availability and Integration

Cloud Deployment

MedGemma is available through multiple channels:

• Hugging Face: weights after accepting the HAI-DEF terms of use
• Gemini Enterprise Agent Platform (formerly Vertex AI): self-deployment from the Model Garden to your own endpoint in the region you choose, including EU regions; check accelerator availability per region
• Large imaging data: for whole-slide pathology and CT/MRI volumes from Cloud DICOM Store or Cloud Storage, Google offers dedicated serving options in the Model Garden

Complementary Models: MedASR and MedSigLIP

Alongside MedGemma 1.5, Google released MedASR, an open speech-to-text model for medical dictation, available via Hugging Face and the Agent Platform. This combination enables clinical documentation with downstream AI analysis. MedSigLIP maps medical images and text into a shared embedding space, for example for classification and image search.

Regulatory Aspects

EU Regulation for Healthcare AI

• MDR/IVDR: Medical device regulations must be strictly observed
• EU AI Act: Medical AI classified as high-risk system
• Liability: Full responsibility lies with the user, not the model developer
• Validation: Clinical validation required before production deployment

Data Privacy in Healthcare Context

• GDPR Art. 9: Special categories of personal data (health data)
• Self-hosting recommended: Patient data should not leave your infrastructure
• Anonymisation mandatory: Required before any processing with AI models
• Data residency: When using cloud services, select EU regions

Self-Hosting for Maximum Control

For medical applications involving patient data, we expressly recommend self-hosting:

• Complete data control and sovereignty
• No data transmission to external providers
• Fulfilment of stringent data protection requirements
• Independence from cloud provider availability

MedGemma’s open weights enable flexible deployment in your own infrastructure, from on-premise servers to private cloud environments.

Our Recommendation

MedGemma 1.5 4B is the starting point for most medical imaging and document tasks: it handles CT and MRI volumes, pathology slides and lab reports and runs on a single server or offline. For complex text tasks with sufficient compute, MedGemma 27B fits, as the text or multimodal variant depending on your needs.

Both are tools for medical research, clinical education and support systems. However, they must never serve as the sole basis for clinical decisions. Per Google, MedGemma must be validated on your own use case and usually adapted further before production use.

For healthcare organisations, we offer specialised AI consulting - particularly regarding compliant integration under MDR, IVDR, and the EU AI Act.