The accuracy and reliability of 3D steady RANS CFD simulations of wind flow in urban environments can be affected by numerical settings including the turbulence model and the imposed roughness heights. In that regard, various k-ε and k-ω turbulence models and roughness height (ks) values are commonly used when predicting wind flow in urban environments. However, it is insufficiently known to which extent the CFD results may be influenced by these settings when simulating wind flows in complex urban environments with large changes in surface roughness. This is the scope of the present paper, for which wind-tunnel (WT) measurements and CFD simulations were performed on a reduced-scale model (1:300) of a district of Livorno (Italy). Mean wind speed (U), turbulent kinetic energy (k) and turbulence dissipation rate (ε) profiles from WT measurements and CFD simulations were compared at 25 positions and deviations between experimental and numerical results were quantified by three metrics: fractional bias, correlation coefficient and fraction of data within a factor of 1.3. The turbulence model selection had a larger impact compared to the surface roughness selection on U, k and ε values. The best and worst performing turbulence models (e.g. for α = 240° at 0.02 m above the bottom) showed a deviation in terms of correlation (0.89 and 0.61, respectively) of about 0.28. Conversely, the best and worst performing roughness set, (e.g. for α = 240° at 0.02 m above the bottom), showed a deviation in terms of correlation (0.77 and 0.78, respectively) of only 0.01.