This data set consists of monthly long-term mean temperature data (degrees C) for the last glacial maximum (21 ka) downscaled to a 10-km grid of North America. The 10-km data were derived using simulated temperature data from 10 general circulation models (GCMs; CCSM4, CNRM-CM5, COSMOS-ASO, FGOALS-g2, GISS-E2-R, IPSL-CM5A-LR, MIROC-ESM, MPI-ESM-P-OA, MPI-ESM-P-OAC, and MRI-CGCM3) run under the PMIP3/CMIP5 (Paleoclimate Modelling Intercomparison Project phase 3 / Coupled Model Intercomparison Project phase 5) “lgm” and “piControl” experiments. The lgm and piControl data are available from the Earth System Grid - Center for Enabling Technologies (ESG-CET; https://esgf-node.llnl.gov/projects/esgf-llnl/). Additional information about the data is available from the CMIP5 (https://pcmdi.llnl.gov/mips/cmip5/) and PMIP3 (https://pmip3.lsce.ipsl.fr/) web sites. The names of the lgm and piControl files we used are listed in the “source_file” global attribute of each GCM temperature netCDF file in this data release. For each GCM, the PMIP3/CMIP5 lgm temperature data were bias corrected using long-term mean differences calculated as the lgm long-term mean minus the piControl long-term mean. These long-term mean differences were regridded to a North America 10-km Lambert azimuthal equal-area grid using the CDO (Climate Data Operators, https://code.mpimet.mpg.de/projects/cdo) bilinear interpolation function “remapbil”. We used ICE-5G (VM2) data (Peltier, 2004, https://doi.org/10.1146/annurev.earth.32.082503.144359) to identify grid cells with ice cover at 21 ka. The interpolated long-term mean differences were applied to CRU CL 2.0 (1961-1990 30-year mean) climate data (New et al., 2002, https://doi.org/10.3354/cr021001). The CRU CL 2.0 data were also regridded to the 10-km grid using local lapse-rate adjusted interpolation (Praskievicz and Bartlein, 2014, https://doi.org/10.1016/j.jhydrol.2014.06.017). The ensemble mean data were calculated using the bias corrected temperature data from each of the 10 GCM simulations.