The strong winds prevalent in high altitude and arctic environments heavily redistribute the snow cover, causing a small-scale pattern of highly variable snow depths. This has profound implications for the ground thermal regime, resulting in highly variable near-surface ground temperatures on the metre scale. Due to asymmetric snow distributions combined with the nonlinear insulating effect of snow, the spatial average ground temperature in a 1 km2 area cannot be determined based on the average snow cover for that area. Land surface or permafrost models employing a coarsely classified average snow depth will therefore not yield a realistic representation of ground temperatures. In this study we employ statistically derived snow distributions within 1 km2 grid cells as input to a regional permafrost model in order to represent sub-grid variability of ground temperatures. This improves the representation of both the average and the total range of ground temperatures. The model reproduces observed sub-grid ground temperature variations of up to 6 °C, and 98 % of borehole observations match the modelled temperature range. The mean modelled temperature of the grid cell reproduces the observations with an accuracy of 1.5 °C or better. The observed sub-grid variations in ground surface temperatures from two field sites are very well reproduced, with estimated fractions of sub-zero mean annual ground surface temperatures within ±10 %. We also find that snow distributions within areas of 1 km2 in Norwegian mountain environments are closer to a gamma than to a lognormal theoretical distribution. The modelled permafrost distribution seems to be more sensitive to the choice of distribution function than to the fine-tuning of the coefficient of variation. When incorporating the small-scale variation of snow, the modelled total permafrost area of mainland Norway is nearly twice as large compared to the area obtained with grid-cell average snow depths without a sub-grid approach.