Turbulence modeling and validation by experiments are key issues in the simulation of micro-scale atmospheric dispersion. This study evaluates the performance of two different modeling approaches (RANS standard k-?? and LES) applied to pollutant dispersion in an actual urban environment: downtown Montreal. The focus of the study is on near-field dispersion, i.e. both on the prediction of pollutant concentrations in the surrounding streets (for pedestrian outdoor air quality) and on building surfaces (for ventilation system inlets and indoor air quality). The high-resolution CFD simulations are performed for neutral atmospheric conditions and are validated by detailed wind-tunnel experiments. A suitable resolution of the computational grid is determined by grid-sensitivity analysis. It is shown that the performance of the standard k-e model strongly depends on the turbulent Schmidt number, whose optimum value is case-dependent and a priori unknown. In contrast, LES with the dynamic subgrid-scale model shows a better performance without requiring any parameter input to solve the dispersion equation.