The aim of the paper is to investigate the ability of the model SWAP-2 (Soil–Water–Atmosphere–Plants) to simulate major components of water balance on a large river basin scale, and to reveal the potentials for model's improvement. The model treats heterogeneity of a large basin explicitly—by means of dividing the basin into a number of computational units (1°×1° grid cells) provided with deterministic effective values for land-surface parameters and atmospheric forcings. These grid cells are linked together by a river network. The components of heat and water balances are simulated separately for each grid cell by a physically based land surface parameterisation scheme. Further, grid simulated runoff is transformed into streamflow by routing models.Different versions of SWAP-2 were validated against naturalised streamflow from 15 drainage basins located within the Red-Arkansas River basin for the period of 1979–1988, and intercompared to get a better insight into the model's performance. The accuracy of the best version was found to be close to the estimated maximum accuracy under the chosen schematisation of drainage basins and prescribed effective input data. Thus, the error of modelled daily streamflow is equal to 0.26mm/day, compared to the estimated minimum error of 0.21mm/day, and the error of modelled annual streamflow is 15mm/year, compared to the estimated minimum error of 12mm/year.Validation of the best version showed that SWAP-2 does not incorporate significant systematic error into the results. As such, the model can operate at a regional scale with satisfactory accuracy under appropriate discretisation of a basin. In this case adjustment of land surface parameters by means of calibration is not necessary, because a large number of computational units greatly reduces the impact of random errors in effective parameters and atmospheric forcings on the simulation of such integral characteristics as streamflow and evapotranspiration from a large basin. When the number of computational units decreases, the relative error of simulations grows up. This enhances the requirements to the accuracy of parameter estimation.