Using the observed mean annual runoff for 1986-1995 from 150 large basins globally, we evaluate the performance of the 14 global land surface models (LSMs) and six Budyko-type hydrological models that are forced by the meteorological data from the second phase of the Global Soil Wetness Project (GSWP-2). The results show that both the 14 LSMs and six Budyko-type models can explain 55-70% of the spatial variations of mean annual runoff across the selected 150 basins. However, the 14 LSMs show larger biases in the simulated mean annual runoff than the Budyko-type models. The LSMs biases are caused by errors in forcing data, model structure and model parameterisation. The errors in the precipitation forcing data are found to be the main cause for biases in the simulated mean annual runoffs by the Budyko-types models, and most likely for biases in the 14 global land surface models too. The GSWP-2 precipitation is noticeably overestimated at Northern high-latitudes, which causes large positive biases for the LSMs in simulating mean annual runoff in these regions. The most LSMs show large biases in the regions with low mean annual precipitation. Underestimation of the GSWP-2 precipitation in Amazon and Orinoco basins results in significant underestimation in the simulated mean annual runoff by all LSMs and Budyko-type models for these regions. The LSMs with smaller biases generally show larger baseflow ratio in wet basins than in dry basins while the LSMs with larger biases generally show smaller baseflow ratio in wet basins than in dry basins. This indicates that errors in model structure can result in large biases in the simulated runoff. Several parameter sensitivity experiments for one LSM are carried out to investigate impacts on simulated mean runoff. The result indicates that +/-20% changes in five key model parameters have relatively smaller impacts on the simulated mean annual runoff across the 150 basins, compared to errors in model structure.