Virchow2G Mini is a distilled, lightweight vision transformer model derived from Virchow2G, designed to deliver high-performance pathology insights with exceptional computational efficiency. Trained on 3.1 million whole slide histopathology images, it serves as a tile-level feature extractor (frozen or finetuned) suitable for a wide range of downstream computational pathology applications. It supports both hematoxylin and eosin (H&E) and immunohistochemistry (IHC) stained slides, enhancing its versatility across various pathology tasks. Despite its compact size, Virchow2G Mini achieves performance comparable to larger models, making it ideal for high-throughput workflows and resource-constrained environments. Virchow2G Mini is based on a ViT-S/14 architecture with 22 million parameters. It processes input images at a native size of 224×224 pixels with a 14×14 patch size. While it omits some of the advanced architectural components of its predecessor—such as SwiGLU activations, LayerScale stabilization, and register tokens—to maximize inference efficiency, it retains robust representational capabilities. The model was distilled from Virchow2G using a modified DINOv2 self-supervised objective, replacing the original KoLeo regularizer with a kernel density estimator and adopting an extended context translation augmentation strategy. Training involved sampling tiles across four magnifications (5x, 10x, 20x, and 40x) and utilized mixed precision (fp16) to optimize efficiency while preserving accuracy. Virchow2G Mini is designed to support flexible downstream fine-tuning or prompting for varied pathology tasks.
