The quality of surface winds derived from four meteorological models is assessed in the semi-enclosed Adriatic Sea over a 2-month period: a global hydrostatic model ECMWF T511 (40 km resolution), a hydrostatic limited area model LAMBO (20 km), and two non-hydrostatic limited area models: LAMI (7 km) and COAMPS™ (4 km). These wind models are used to drive a 2 km resolution wave model (SWAN) of the Adriatic, and wind and wave results are compared with observations at the ISMAR oceanographic tower off Venice. Waves are also compared at buoy locations near Ancona and Ortona. Consistently with earlier studies, the ECMWF fields underestimate the wind magnitude and do not reproduce the known spatial structure of strong wind events. The results show that the higher-resolution, limited area models LAMI and COAMPS exhibit better amplitude response than the coarser ECMWF: there is a 3- to 4-fold reduction of the wind underestimation at the platform (from 36% to 8–11%). The wave response is also improved with LAMI and COAMPS: there is a 2-fold reduction in the underestimation of wave heights at the platform. These non-hydrostatic models also produce wind fields with more realistic small-scale, spatial structure during strong wind events. The temporal correlation between observed and modelled wind, however, is highest with the global ECMWF model due to the fact that large-scale features can be predicted deterministically, whereas small-scale features can only be predicted stochastically. Models with less small-scale structure have better correlation because they have less “noise.” This explanation is supported by increased correlation between modelled and observed waves, the waves representing a smoothing of the wind over fetch and duration. Although there is room for improvement, the high-resolution, non-hydrostatic models (LAMI and COAMPS) offer significant advantages for driving oceanographic simulations in semi-enclosed basins such as the Adriatic Sea.