Precipitation forecasting experiments have been carried out with HIRLAM, a numerical weather prediction model. The model has been run with three different horizontal gridlengths: 22, 11 and 5.5 km. An attempt has been made to estimate the appropriate magnitude of the horizontal diffusion, which is used to control small-scale noise in the model, by looking at kinetic energy spectra. It is shown that, for the higher resolutions, smoothing the orography gives smoother precipitation patterns. The small-scale precipitation resulting from runs with the original orography has negligible extra skill compared to the smoothed orography runs. The results show that the model is able to forecast good precipitation amounts, even with 22 km gridlengths. No significant improvements occur when the higher horizontal resolutions are used. Experiments have been performed using the tendencies of the physical parameterisations computed on a coarser grid than that of the dynamics. The resulting precipitation patterns are very similar and this indicates a more economical way of integrating the model.