The inter-annual variability and potential predictability of 850 hPa temperature (T850), 500 hPa geopotential (<phi>500) and 300 hPa stream function (<psi>300) simulated by the models participating in the Atmospheric Model Intercomparison Project (AMIP) are examined. The total inter-annual variability is partitioned into a potentially predictable component which arises from the forcing implied by the prescribed SST and sea-ice evolution, or from sources internal to the simulated climate, and an unpredictable low frequency component induced by "weather noise". There is wide variation in the ability to simulate observed inter-annual variability, both total and weather-noise induced. A majority of models under simulate seasonal mean <phi>500 variability in DJF and JJA and over simulate <psi>300 variability in JJA. All but one model simulates less T850 total inter-annual variability than in the analysed data. There is little apparent connection between gross model characteristics and the corresponding ability to simulate observed variability, with the possible exceptions of resolution.