The validity of a radiative transfer model can be checked either by comparing its results with measurements or with solutions for artificial cases. Unfortunately, neither type of comparison can guarantee that the spectral UV surface irradiance is accurately calculated for real atmospheric cases. There is a need therefore for benchmarks, i.e., standard results that can be used as a validation tool for UV radiation models. In this paper we give such benchmarks for six cloud-free situations. The chosen cases are characterized by different values of solar zenith angle, ozone column, aerosol loading, and surface albedo. Observations are also available for these cases to allow a further comparison between model results and measurements. An intercomparison of 12 numerical models is used to construct the benchmarks. Each model is supplied with identical input data, and a distinction is made between models that assume a plane-parallel geometry and those that use a pseudospherical approximation. Differences remain between the model results, because of different treatments of the input data set. Calculations of direct and global transmission and direct and global irradiance are within 3% for wavelengths longer than 320 nm. For the low-Sun cases the calculations are within 10% for wavelengths longer than 300 nm. On the basis of these calculations, six benchmark UV spectra (295-400 nm) are established with a standard deviation of 2%. Relative standard deviations are higher for the lowest absolute intensities at low Sun (5% at 300 nm). The variation between models is typically less than the variation seen between model and measurement. Differences between the benchmarks and the observed spectra are mainly due to the uncertainty in the input parameters. In four of the six cases the benchmarks agree with the observed spectra within 13% over the whole UV spectral region.