Google Unveils MedGemma 27B and MedSigLIP: Open-Source Breakthroughs in Multimodal Medical AI

Introducing MedGemma 27B Multimodal and MedSigLIP

Google DeepMind and Google Research have launched two innovative open-source models for medical AI under the MedGemma framework: MedGemma 27B Multimodal, a large-scale vision-language foundation model, and MedSigLIP, a lightweight image-text encoder tailored for healthcare. These represent the most advanced open-weight models released within the Health AI Developer Foundations (HAI-DEF).

MedGemma Architecture and Purpose

MedGemma expands on the Gemma 3 transformer backbone, adapting it for healthcare challenges including diverse data types, limited supervision, and practical deployment needs. It processes both clinical text and medical images, supporting complex tasks such as diagnosis, report generation, retrieval, and agentic reasoning.

MedGemma 27B Multimodal: Advanced Multimodal Reasoning

This 27-billion parameter transformer decoder model accepts interleaved medical images and text, powered by a high-resolution (896×896) image encoder based on SigLIP-400M trained on over 33 million medical image-text pairs from various specialties like radiology and histopathology.

Key features include:

87.7% accuracy on MedQA (text-only version), surpassing all open models under 50B parameters.
Effective multi-step decision-making in simulated diagnostic environments (AgentClinic).
End-to-end reasoning across patient history, images, and genomics for personalized treatment.

Clinical applications include multimodal question answering, radiology report generation, cross-modal retrieval, and clinical agent simulation.

MedSigLIP: Lightweight and Efficient Image-Text Encoder

MedSigLIP, with 400 million parameters and a resolution of 448×448, supports edge and mobile deployment while excelling in zero-shot and linear-probe classification tasks across dermatology, ophthalmology, histopathology, and radiology.

Its benchmarks include:

Outperforming ELIXR-based chest X-ray models by 2% AUC.
Achieving 0.881 AUC on dermatology classification over 79 conditions.
Delivering 0.857 AUC on diabetic retinopathy classification.
Matching or exceeding state-of-the-art in cancer subtype histopathology classification.

The model uses cosine similarity between image and text embeddings for zero-shot tasks, with optional lightweight fine-tuning via logistic regression.

Open Source and Deployment

Both MedGemma 27B and MedSigLIP are fully open source, with available weights, training scripts, and tutorials. They integrate seamlessly with the Gemma ecosystem and can be deployed on single GPUs, including mobile hardware via quantization and distillation, requiring minimal code.