The semi-distributed hydrological model SWAT was built for the Yarlung Zangbo River Basin. After calibration (1989-1998) and validation (1999-2005), the monthly runoff was simulated at three hydrological stations (Nugesha, Yangcun, Nuxia). We selected two sets of parameters to analyze the performance of SWAT model through Nash-Sutcliffe Efficiency (NS), determination coefficient ( R2 ) and relative error (PBIAS). The results show that for Parameter Set 1, R2 at three stations were as high as 0.9, NS were 0.9, 0.92 and 0.72, PBIAS were 7.1%,18.5% and 34.2% respectively during the calibration. However, the simulations of runoff components are unrealistic. For Parameter Set 2, although the evaluation indexes were not very high, its runoff components are reasonable. This is the same case during the validation. Therefore, it is recommended that the simulation skills of runoff components be considered during model calibration and validation.