The performance of three large-eddy simulation (LES) models in predicting the evolution of a shear layer at moderate Reynolds number in a linearly stratified background is investigated. Results of the Dynamic Smagorinsky, Ducros, and Wall-Adapting Local Eddy Viscosity (WALE) models are compared against those of direct numerical simulation (DNS). Two levels of grid refinement are employed to assess the change in the models’ capabilities with varying resolution. Of particular interest is the ability of the models to capture the evolution of instabilities as well as accurately quantify turbulence statistics. Evolution of momentum thickness, local buoyancy flux, shear, and gradient Richardson number show good agreement of the LES models with the DNS. A comparison of the turbulent kinetic energy (TKE) and its budget indicates good capture of turbulence evolution in the LES models. There is a moderate over prediction of spatially integrated TKE during a short period when the integrated TKE is at its maximum. This feature is traced to the underestimate of turbulent dissipation rate in the LES. Coarsening grid resolution significantly increases the discrepancy in dissipation between the WALE model and the DNS while defects due to the coarser grid resolution are mild in the case of the Ducros model. All of the LES models overestimate the peak values of eddy viscosity and eddy diffusivity although the Ducros model produces the closest agreement with the DNS. Furthermore, the Ducros model is found to require fewer CPU hours than the DNS or other LES models.