Streamline pathology foundation model by cross-magnification distillation

Foundation models (FM) have transformed computational pathology but remain computationally prohibitive for clinical deployment due to their massive parameter counts and high-magnification processing requirements. Here, we introduce XMAG, a lightweight FM developed through cross-magnification distillation that transfers knowledge from a state-of-the-art 20x magnification teacher to an efficient 5x magnification student architecture. XMAG employs a compact backbone and operates entirely at 5x, requiring 11.3 times fewer patches per whole slide image (WSI) compared to existing approaches. Our novel distillation framework incorporates dual-level knowledge transfer, aligning both global image representations and local spatial features across magnification levels through carefully designed projection heads and spatial token mapping. We trained XMAG on 3.49 million images curated from publicly available datasets and evaluated performance across seven clinically relevant histopathology analysis tasks spanning multiple cancer types. Across four WSI-level classification tasks, XMAG achieved an average AUC within 1% of substantially larger foundation models while delivering up to 29.5-fold faster encoder-level processing, corresponding to an encoder throughput of 8.8 WSIs per minute in a controlled patch-encoding benchmark. We also found that performance differences remained task-dependent, with larger gaps observed for fine-grained tasks such as ISUP grading and patch-level phenotyping. We further explore end-to-end XMAG training as a direction for task-specific adaptation of the compact model. These results suggest that cross-magnification distillation is a promising strategy for developing efficient low-magnification pathology foundation models and may support more scalable WSI analysis in computationally constrained research and translational settings.