Abstract A modified semi-Lagrangian scheme is proposed in the context of semi-implicit forecast models to reduce the important distortion of topographically forced waves that is produced when the Courant–Friedrichs–Lewy (CFL) number is greater than 1. The improved semi-Lagrangian scheme combines the original semi-implicit formulation and the spatial averaging of all nonlinear terms. The impact of the spatial averaging is assessed in two baroclinic forecast models: a global spectral model and a regional gridpoint model. The modified semi-implicit semi-Lagrangian scheme is shown to improve short- and medium-range forecasts, and to increase the efficiency of the models by reducing the number of interpolations by 20%–40%.