Earth System Model (ESM)-derived Surface Mass Balance (SMB) serves as input for ice sheet models and to make estimates of sea level rise. However, there are few previous studies examining the biases in ESMs when modelling SMB under the historical (1985–2014) period, which may elucidate biases in future projections. Regional Climate Models (RCMs) are often used as a computationally expensive step in improving modelled SMB at the present day and for future projections. It is also unclear whether they add value to SMB estimates on the Antarctic Ice Sheet (AIS), where the topography of the Plateau is relatively smooth. We evaluate SMB from 4 ESMs, 13 ESM-driven RCM simulations, 3 reanalyses and 13 reanalysis-driven RCM simulations, with 76003 AIS SMB observations compiled in the SUMup dataset. We find that ESMs can have consistent Antarctic-wide biases that are related to their temperature biases, and regional biases specific to each ESM. Downscaling with RCMs greatly improves the modelling of SMB relative to ESMs: they systematically increase the correlation coefficient (R; from 0.71 for ESMs to 0.86 for ESM-driven RCMs) and Perkins Skill Score (PSS; from 0.38 to 0.67), and reduce the Root-Mean-Squared Error (RMSE; from 17.90 cm to 15.60 cm). However, RCM downscaling does not systematically outperform reanalyses according to the same metrics (R from 0.88 to 0.90; PSS from 0.58 to 0.69; and RMSE from 11.45 cm to 11.22 cm). When downscaling, RCMs carry strong biases from their driving models, and performance improvement often stems from regional rather than continent-wide improvements. An important caution in interpreting our results is that observations are unevenly distributed in both time and space, hindering a truly unbiased ice sheet-wide assessment of SMB. However, we find that the density of observations is sufficient to identify the biases between models and observations in Dronning Maud Land and West Antarctica, but there are insufficient measurements in the rest of East Antarctica and the Peninsula, where high added value of RCMs is expected, suggesting that these regions should be a target for future observational campaigns.