A comparison of 8 regional atmospheric model systems was carried out for a three-month late summer/early autumn period in 1995 over the Baltic Sea and its catchment area. All models were configured on a common grid using similar surface and lateral boundary conditions, and ran in either data assimilation mode (short term forecasts plus data assimilation), forecast made (short term forecasts initialised daily with analyses) or climate mode (no re-initialisation of model interior during entire simulation period). Model results presented in this paper were generally post processed as daily averaged quantities, separate for land and sea areas when relevant. Post processed output was compared against available analyses or observations of cloud cover, precipitation, vertically integrated atmospheric specific humidity, runoff, surface radiation and near surface synoptic observations. The definition of a common grid and lateral forcing resulted in a high degree of agreement among the participating model results for most cases. Models operated in climate mode generally displayed slightly larger deviations from the observations than the data assimilation or forecast mode integration, but in all cases synoptic events were well captured. Correspondence to near surface synoptic quantities was good. Significant disagreement between model results was shown in particular for cloud cover and the radiative properties, average precipitation and runoff. Problems with choosing appropriate initial soil moisture conditions from a common initial soil moisture field resulted in a wide range of evaporation and sensible heat flux values during the first few weeks of the simulations, but better agreement was shown at later times.