A successful hydrological seasonal forecast needs skillful seasonal climate predictions from coupled atmosphere-ocean general circulation models (CGCMs), advanced downscaling methods that resolve scale issues and correct biases, and accurate hydrologic initial conditions. Recently, NCEP has transitioned operationally to the second generation of their CGCM, the Climate Forecast System version 2 (CFSv2), with advanced physics, increased resolution and refined initialization to improve the seasonal climate forecasts. Through deterministic assessment and probabilistic evaluation, we demonstrate the advantage of CFSv2 in predicting land surface air temperature and precipitation as compared with CFSv1 and European seasonal forecast models. In order to drive the Variable Infiltration Capacity (VIC) land surface model to provide seasonal forecasts of drought and flood with lead times of up to six months, the CFSv2 forecasts have been bias-corrected and downscaled to 1/8° over CONUS by an optimal Bayesian probabilistic approach. The downscaling method has different prior distributions for below-normal, normal and above normal years; and estimates the likelihood function by using conditional distribution of the optimal ensemble mean given the observation, as well as the conditional distribution of the ensembles given their optimal ensemble mean values that are calibrated against observation. The proposed optimal Bayesian approach shows significant advantage in downscaling precipitation in severe pluvial or dry years against traditional Bayesian method. A 62-year (1949-2010) VIC offline simulation driven by the merged UW and NLDAS2 data has been conducted to provide initial conditions and hydrologic climatology, and comprehensive ensemble hydrological reforecasts during 1982-2009 has been carried out over CONUS. The predicted streamflow and soil moisture are respectively compared with observed streamflow over 14 large river basins and modeled soil moisture from offline simulation.