A wide range of numerical wind flow models are available to simulate atmospheric flows. For wind resource mapping, the traditional approach has been to rely on linear Jackson–Hunt type wind flow models. Mesoscale numerical weather prediction (NWP) models coupled to linear wind flow models have been in use since the end of the 1990s. In the last few years, computational fluid dynamics (CFD) methods, in particular Reynolds-averaged Navier–Stokes (RANS) models, have entered the mainstream, whereas more advanced CFD models such as large-eddy simulations (LES) have been explored in research but remain computationally intensive. The present study aims to evaluate the ability of four numerical models to predict the variation in mean wind speed across sites with a wide range of terrain complexities, surface characteristics and wind climates. The four are (1) Jackson–Hunt type model, (2) CFD/RANS model, (3) coupled NWP and mass-consistent model and (4) coupled NWP and LES model. The wind flow model predictions are compared against high-quality observations from a total of 26 meteorological masts in four project areas. The coupled NWP model and NWP-LES model produced the lowest root mean square error (RMSE) as measured between the predicted and observed mean wind speeds. The RMSE for the linear Jackson-Hunt type model was 29% greater than the coupled NWP models and for the RANS model 58% greater than the coupled NWP models. The key advantage of the coupled NWP models appears to be their ability to simulate the unsteadiness of the flow as well as phenomena due to atmospheric stability and other thermal effects. Copyright © 2012 John Wiley & Sons, Ltd.