This work proposes HURST, a heterogeneity-adaptive urban foundation model for spatiotemporal prediction, designed to address the spatial heterogeneity widely present in urban spatiotemporal data. HURST automatically identifies regions with distinct spatial patterns through a self-partitional mixture-of-spatial-experts network and employs an error-guided adaptive spatiotemporal masking strategy to dynamically adjust the training process across regions, thereby improving the model’s ability to capture heterogeneous urban patterns and generalize across different settings.