The aim of this study is to develop and test a methodology to explicitly resolve glacier energy and mass balance in regional climate models (RCMs). There is increasing interest in the ability to represent mountain glaciers in climate models, to enable either one-way forced or fully-coupled simulations of glacier response to climate change. However, mountain glaciers are generally subgrid features (areas of less than 10 km2) occupying high elevations in complex mountain terrain, and it is important to examine whether RCM-derived meteorological fields in mountain regions are accurate enough to model glacier mass balance. We use downscaled RCM data to force a glacier model at five different sites in western North America, where long-term glacier mass balance observations allow us to evaluate different glacier modelling and climate downscaling strategies. The model runs on a pre-processed, gridded database of 100-m resolution topographic characteristics. Our reference glacier simulations use this high resolution grid to explicitly simulate glacier surface and energy balance, but this is computationally demanding. We test different mosaic approaches for subgrid characterization of the glaciers, where glaciers in an RCM grid cell are distributed over different terrain classes, similar to hydrological response units. The subgrid mosaic approach drastically increases computational efficiency without affecting mass balance simulations for the five glaciers by more than a few percent. Elevation distribution within a grid cell is the most important element of the subgrid terrain characterization; subgrid elevation bands need to be resolved to 200 m or less in order to keep seasonal mass balance biases lower than 2%. Good simulations of glacier mass balance can be achieved with glacier-specific tuning, with correlation coefficients of ~ 0.8 between modelled and measured annual glacier mass balances over the period 1995–2014. However, there are large and regionally-variable biases in the modelled meteorological fields, and we cannot achieve good results without local bias correction.