We compare measurements of the turbulent and radiative surface energy fluxes from an automatic weather station (AWS) on Larsen C Ice Shelf, Antarctica with corresponding fluxes from three high-resolution atmospheric models over a one-month period during austral summer. All three models produce a reasonable simulation of the (relatively small) turbulent energy fluxes at the AWS site. However, biases in the modelled radiative fluxes, which dominate the surface energy budget, are significant. There is a significant positive bias in net shortwave radiation in all three models, together with a corresponding negative bias in net longwave radiation. In two of the models, the longwave bias only partially offsets the positive shortwave bias, leading to an excessive amount of energy available for heating and melting the surface, while, in the third, the negative longwave bias exceeds the positive shortwave bias, leading to a deficiency in calculated surface melt. Biases in shortwave and longwave radiation are anticorrelated, suggesting that they both result from the models simulating too little cloud (or clouds that are too optically thin). We conclude that, while these models may be able to provide some useful information on surface energy fluxes, absolute values of modelled melt rate are significantly biased and should be used with caution. Efforts to improve model simulation of melt should initially focus on the radiative fluxes and, in particular, on the simulation of the clouds that control these fluxes.