Developing modelling tools that help to assess the spatial distribution of water resources is a key issue to achieve better solutions for the optimal management of water availability among users. Distributed hydrological models which explicitly consider spatial heterogeneities within catchments are consequently pointed as a solution. However, the analysis of previous studies demonstrates that the benefit of a distributed model over a lumped model is not systematic. In this study, we investigate two main cases where a semi-distributed model is expected to outperform a lumped model. The first one concerns local influence of dams, the second concerns simulation at ungauged locations. Firstly, for the case of dam influences, we adapted the lumped approach of Payan et al. (2008) to a semi- distributed framework which uses water storage variation of dams as an additional input data. In this way, the specific location of existing reservoirs inside a catchment is explicitly taken into account. A comparison to the lumped strategy demonstrates better performance of the semi- distributed model at influenced locations, especially for reproducing low-flow characteristics. Secondly, we evaluated the capacity of the model to estimate streamflow at ungauged locations. This is done through a "leave-one-out" evaluation where each gauged catchment is successively considered as ungauged. We evaluated the benefit of a semi-distributed model, which can take advantage of calibrated sub-catchments inside the ungauged one, to outperform a regionalized lumped model. A large database of over 1270 catchments spread all over France (drainage area between 15 km² and 110 000 km²) enables to explore numerous configurations. The semi-distributed rainfall-runoff model used to investigate those two points is based on the lumped GR5J model (de Lavenne et al. 2016): it runs at the daily time step, and has five parameters for each sub-catchment as well as a streamflow velocity parameter for flow routing. Its structure is based on two stores, one for runoff production and one for routing. The calibration of the model is performed from upstream to downstream sub-catchments, which efficiently uses spatially-distributed streamflow measurements and facilitate regionalization procedure. References: de Lavenne, A.; Thirel, G.; Andréassian, V.; Perrin, C. & Ramos, M.-H. (2016), Spatial variability of the parameters of a semi-distributed hydrological model, PIAHS 373, 87-94. Payan, J.-L.; Perrin, C.; Andréassian, V. & Michel, C. (2008), How can man-made water reservoirs be accounted for in a lumped rainfall-runoff model?, Water Resour. Res. 44(3), W03420.