<p>Hydrological models are essential for predicting water flux dynamics, including extremes, and managing water resources, yet traditional process-based large-scale models often struggle with accuracy and process understanding due to their inability to represent complex, non-linear hydrometeorological processes, limiting their effectiveness in local conditions. Here we explore hybrid methods combining process-based modelling and statistical or machine learning post-processors to improve streamflow predictive accuracy, including extremes, across Europe’s hydro-climatic gradient. We investigate various post-processing methods, such as random forest, long short-term memory model, quantile mapping and generalised linear model, demonstrating notable improvements in model performance, in terms of reducing errors in total volume and extremes and increasing robustness across diverse climatic and geographic conditions. We further show that hydrologic similarity is one of the key drivers that control the hybrid approach’s improvements, together with other basin characteristics, such as mean precipitation and mean temperature. Our results also reveal spatial complementarity among the post-processing methods, with no absolute superiority identified from a single method, pointing towards multi-model averaging approaches for the future evolution of hybrid hydrological modelling.</p><p></p>