Model error is a main concern for the EnKF in assimilating real observations. In this study we investigated several methods to handle model errors in EnKF including model bias and system-noise. Model errors were introduced by assimilating observations generated from the NCEP/NCAR reanalysis fields into the SPEEDY model. Our results suggest that the performance of the EnKF without accounting for model errors is seriously degraded compared with that in the perfect model scenario. The pure bias removal methods (DdSM, Dee and da Silva 1998; LDM, Danforth et al. 2007) are not able to beat the multiplicative or additive inflation schemes that account for the effects of total model errors. By contrast, when the bias removal methods (DdSM+ and LDM+) are supplemented by additive noise for representing the system-noise, they outperform the inflation schemes. Of these augmented methods, the LDM+, where the constant bias, diurnal bias and state-dependent errors are estimated from a large sample of 6-hour forecast errors, gives the best results.