Modeling has become a very useful tool in air quality management. The use of an air quality model requires comparison between model results and previous observations in order to determine the capacity of the model to reproduce air pollution episodes. In this paper the influence of three different model resolutions on model predictions has been analyzed over Madrid area for 2004. A lower mean normalized absolute error was found for the highest resolution domain, when comparing hourly-predicted ozone to 2004 observations. The improvement of model predictions is more clearly observed for NO₂. When considering an episode occurred in July, 2004, this improvement in model performance is significantly reduced. For this episode, meteorological evaluation indicates that temperature and speed predictions for the coarsest domain present worse agreement to observations than those for the other two domains.