All mesoscale model forecasts have biases, some quite substantial. Many of these biases are systematic and repeatable, thus offering the potential for their removal in post-processing. Until now, most bias removal has used the Model Output Statistcs (MOS) or perfect prog approaches that are applied only at observation locations. However, there is a need for model bias removal on the model grid itself. The demand for grid-based bias removal has become particularly acute as the National Weather Service, private sector, and other forecast groups move to grid-based forecast dissemination. Model grids, often with substantial biases, are generally the starting points to create this grids, which can be modified by human forecasters. This presentation will describe the evaluation at the University of Washington of two approaches to grid-based bias removal, both using the real-time UW MM5 system run at 36, 12, and 4-km grid spacing. The first method, based on gridded analyses from RUC or EDAS, averages the biases (deviations from analysis) over the previous two weeks on an hourly basis for each forecast cycle. The second method is based directly on observations, using the biases at nearly observation locations with similar land use and elevation to compute the biases (hourly) at each grid point. This presentation will compare the results of both methods over an extended period of time at the surface to determine the value of grid-based bias removal in producing forecasts of increased value to the user community.