Elevated dissolved organic carbon (DOC) concentrations are a major concern for ecosystems and drinking water supply. Data-driven studies revealed variable functioning of different landscape units (upland, riparian zone, and groundwater) in catchment DOC mobilization and export. However, lumped and landscape-explicit (separating upland and riparian zone) model structures are generally calibrated to stream DOC concentrations, while the internal DOC dynamics often do not receive sufficient attention. Here, we developed a flexible model with a lumped and landscape-explicit structure for four headwater catchments in the Harz Mountains, Germany. We evaluated these models under a baseline calibration (only against stream DOC concentration) and a constrained calibration (using stream DOC and internal DOC concentrations). Under the baseline calibration, both model structures reproduced stream DOC dynamics with acceptable performance in some catchments (Kling–Gupta efficiency of behavioural simulations>0.6), but with unrealistically high groundwater DOC. By contrast, the constrained calibration reduces the KGE for stream DOC concentrations but produces internal DOC dynamics consistent with the imposed observational and process-based constraints. Additionally, the landscape-explicit model structure is more robust than the lumped model structure under changing boundary conditions. Our study thus shows that robust catchment DOC modelling requires not only explicit representation of different landscape units, but also calibration constraints based on DOC concentrations within those units.