To improve the analysis quality by incorporating the flow-dependent ensemble covariance into the variational data assimilation system, the new GRAPES (global/regional assimilation and prediction system) hybrid-3Dvar system was built. The new system is based on the GRAPES regional 3DVar system, which uses the statistic covariance, and was built by augmenting the state vectors with another set of control variables preconditioned upon the ensemble dynamic covariance. The new hybrid-3DVar system and the localization method were verified through a single-observation assimilation experiment with ensemble samples produced by the 3D-Var’s control variable perturbation method. The real observation assimilation and forecast experiment for Typhoon Soudelor yielded the following conclusions: (1) The background covariance, which is represented by ensemble samples, is flow-dependent, and the root mean square spread in the ensemble of momentum field and mass field is largest near the typhoon center. (2) The analysis increments of the new hybrid-3DVar have a more detailed structure and more medium- and small-scale information. (3) The analysis and 24 h prediction qualities of model variables in the new hybrid-3DVar are significantly improved compared with the 3DVar system, and the precipitation position predictions are more accurate. (4) The 24 h forecast track of Typhoon Soudelor is closer to the observational one, and the 48 h-predicted intensity also approaches the real observation.