Paige Foundation Models (Virchow & PRISM)

Paige Foundation Models — comprising the Virchow vision AI suite and PRISM multi-modal AI — are purpose-built computational pathology models developed by Paige, now a Tempus company. Trained on the world's largest dataset of digitized pathology slides, these foundation models are designed for researchers, AI developers, and clinical teams seeking to accelerate discovery, improve diagnostic precision, and build next-generation pathology AI applications at scale.

The two model suites address distinct but complementary needs: Virchow provides powerful vision-based feature extraction across tissue types and staining modalities, while PRISM combines large vision models with large language models to deliver slide-level analysis and automated clinical report generation. Together, they offer a comprehensive platform for building and deploying advanced pathology AI.

Virchow Vision AI Suite

• Virchow: The first million-slide foundation model for cancer, trained on 1.5 million H&E whole-slide images (WSIs), covering 17 tissue types at 20x magnification with 632 million parameters. Published in Nature Medicine and available on Hugging Face and Azure AI Foundry.
• Virchow2: An expanded model trained on 3 million H&E and 400,000 IHC WSIs, covering 40 tissue types across 5x, 10x, 20x, and 40x magnifications with 632 million parameters. Features multi-scale training for boosted precision. Available on Hugging Face, Azure AI Foundry, and via commercial license.
• Virchow2G: Paige's largest model to date, with 1.8 billion parameters, trained on the same 3 million H&E and 400,000 IHC WSI dataset across 40 tissue types and four magnification levels. Optimized for high-stakes pathology AI applications. Available via commercial license only.
• Virchow2G-Mini: A lightweight alternative with 21.6 million parameters, trained on the same dataset as Virchow2G, designed to deliver high-throughput performance on limited compute resources. Available via commercial license only.

PRISM Multi-Modal AI

• Combines Large Vision Models (LVMs) with Large Language Models (LLMs) to enable slide-level analysis and insight extraction.
• Built using Virchow as its vision backbone, PRISM analyzes whole-slide images at the slide level rather than patch level.
• Trained on 587,000 H&E WSIs across 16 tissue types at 20x magnification, alongside 195,000 clinical text reports.
• Automated Clinical Report Generation: Converts whole-slide images into free-text pathology reports, including cancer detection, subtyping, and biomarker predictions.
• Zero-Shot Cancer Classification: Detects cancer subtypes and predicts biomarkers without requiring extensive labeled training data.
• Multi-Modal Intelligence: Generates text-based insights by jointly learning from imaging and clinical report data.
• Available on Hugging Face and Azure AI Foundry.

Key Capabilities and Benefits

• Faster AI Model Development: Reduces the time and cost associated with training pathology AI models from scratch by providing pre-trained, high-performance foundation models.
• Regulatory-Grade Performance: Engineered with clinical excellence and compliance in mind, supporting use in high-stakes diagnostic and research settings.
• Flexible Integration: Compatible with leading digital pathology platforms, with availability across Hugging Face, Azure AI Foundry, and commercial licensing options depending on the model tier.

Paige Foundation Models are accessible through multiple deployment pathways — including open access via Hugging Face, cloud-based deployment through Azure AI Foundry, and commercial licensing for enterprise and regulated environments — making them suitable for a broad range of computational pathology workflows from academic research to clinical AI development.