Accurate and reliable rainfall-runoff modeling is essential for flood forecasting, drought monitoring, and reservoir management. Currently, the benchmark practice in data-driven modeling is to train Long Short-Term Memory (LSTM) models across multiple catchments to learn diverse hydrological responses. However, the recurrent structure and inductive bias of LSTMs tend to emphasize recent inputs, making them less effective at capturing very long-range dependencies. Their sequential, thus not parallel structure also limits scalability for very large-scale datasets. With the rapid growth of Earth observations, a key challenge still remains whether a model can effectively train data from multi-sources, learn long-range temporal dependencies, and scale performance with increasing data availability? In this study, we introduce an Attention-based model that outperforms LSTM with a median Nash-Sutcliffe Efficiency (NSE) of 0.781 and Kling-Gupta Efficiency (KGE) of 0.808 compared to LSTM’s NSE of 0.765 and KGE of 0.760, respectively, without relying on ensemble runs or multiple precipitation datasets. The model also demonstrates improved skills in high-flow events and shows continued performance with additional training data. In addition, the Attention-based model offers enhanced interpretability, by assigning dynamic importance to historical meteorological forcings and captures diverse rainfall–runoff responses. Furthermore, by embedding catchment-specific attributes, the model adapts to different hydrological conditions and allows extraction of generalized knowledge through activation-based analysis. This study highlights the potential of Attention-based models in hydrological modeling, particularly for scalable, data-driven flood prediction at continental to global scales.