Spatially distributed meteorological information at the slope scale is relevant for many processes in complex terrain, yet information at this sub-km spatial resolution is difficult to obtain. While downscaling to kilometer resolutions is well described in literature, moving beyond the kilometer scale is not. In this work, we present a methodical comparison of three downscaling methods of varying complexity, that are used to downscale data from the Numerical Weather Prediction model COSMO-1 at 1.1 km horizontal resolution to 250 and 50 m over a domain of highly complex terrain in the Swiss Alps. We compare WRF, a dynamical atmospheric model; ICAR, a model of intermediate complexity; and TopoSCALE, an efficient topography-based downscaling scheme. Point-scale comparisons show similar results amongst all three models w.r.t. mean-error statistics, but underlying dynamics are different. Ridge-flow interactions show reasonable agreement between WRF and ICAR at 250 m model resolution. However, at 50 m resolution WRF is able to simulate complex flow patterns that ICAR cannot. Validation against Lidar data suggests that only WRF is able to capture preferential deposition of snow. Based on these findings and the significant reduction in computational costs, ICAR is a cost efficient alternative to WRF at the 250 m resolution. TopoScale performs very well in point-scale comparisons, but it is unclear if this can be attributed to the model itself or to the forcing data and the observations assimilated therein. Further study is required to quantify this effect.