To advance the discovery of scale-relevant hydrological laws while better exploiting massive multi-source data, merging machine learning into process-based modeling is compelling, as recently demonstrated in lumped hydrological modeling. This article introduces MLPM-PR, a new and powerful framework standing for Multiscale spatially distributed Learnable Physical Modeling and learnable Parameter Regionalization with data assimilation. MLPM-PR crucially builds on a differentiable model that couples (i) two neural networks for processes learning and parameters regionalization, (ii) grid-based conceptual hydrological operators, and (iii) a simple kinematic wave routing. The approach is tested on a challenging flash flood-prone multi-catchment modeling setup at high spatio-temporal resolution (1km, 1h). Discharge prediction performances highlight the accuracy and robustness of MLPM-PR compared to classical approaches in both spatial and temporal validation. The physical interpretability of spatially distributed parameters and internal states shows the nuanced behavior of the hybrid model and its adaptability to diverse hydrological responses.