The focus of this study is on the validation of simulated streamflow from the four North American Land Data Assimilation System (NLDAS) land surface models (Noah, Mosaic, Sacramento, VIC) and their ensemble mean. Comparisons are made with the 28-year (1 October 1979 30 September 2007) U.S. Geological Survey (USGS) observed streamflow (from day to year) for 961 small basins and 8 major basins over the continental United States (CONUS). The relative bias, anomaly correlation and Nash-Sutcliffe efficiency for different time scales (from daily to annual) are used to assess model-simulated streamflow. The Sacramento and VIC models simulate the mean annual runoff and evapotranspiration well when compared with the observations, while the Noah (the Mosaic) model overestimates (underestimates) mean annual runoff and underestimates (overestimates) mean annual evapotranspiration. The ensemble mean is closer to the mean annual observed streamflow for both 961 small basins and 8 major basins than the mean from any individual model. All of the models, as well as the ensemble mean, have large daily, weekly, monthly, and annual streamflow anomaly correlations for most basins over the CONUS, implying strong simulation skill. However, the daily, weekly and monthly Nash-Sutcliffe efficiency (NSE) analysis results are not necessarily encouraging, in particular for daily streamflow. The Noah and Mosaic model are useful (NSE > 0.4) only for about 10% of all 961 small basins, the Sacramento and VIC model are useful for about 30% of 961 small basins, and ensemble mean is useful for about 42% of 961 small basins. As the time scale increases, the NSE increases as expected. However, even for monthly streamflow, the ensemble mean is useful for about 75% of 961 small basins.