Accurate streamflow simulations in large river basins are crucial to predict timing and magnitude of floods and droughts and to assess the hydrological impacts of climate change. Water balance models have been used frequently for these purposes. Compared to water balance models, however, land surface models carry the potential to more accurately estimate hydrological partitioning and thus streamflow, because they solve the coupled water and energy balance and are able to exploit a larger part of the information provided by regional climate model output than water balance models. Owing to increased model complexity, however, they are also more difficult to parameterize. The purpose of this study is to investigate and compare the accuracy of streamflow simulations of a water balance approach (Spatial Tools for River basins and Environment and Analysis of Management (STREAM)) and a land surface model (Variable Infiltration Capacity (VIC)) approach. Both models are applied to the Rhine river basin using regional climate model output as atmospheric forcing, and are evaluated using observed streamflow and lysimeter data. We find that VIC is more robust and less dependent on model calibration. Although STREAM performs better during the calibration period (Nash-Sutcliffe efficiency (E) of 0.47 versus E = 0.29 for VIC), VIC more accurately simulates discharge during the validation period, including peak flows (E = 0.31 versus E = 0.21 for STREAM). This is the case for most locations throughout the basin, except for the Alpine part where both models have difficulties due to the complex terrain and surface reservoirs. In addition, the annual evaporation cycle at the lysimeters is more realistically simulated by VIC. Copyright 2008 by the American Geophysical Union.