If the goal is to recommend the most reliable genotypes then priority should be given to selecting the most accurate model. The additive main effect and multiplicative interaction (AMMI) model has often been shown to be more accurate than the cell means model (means averaged over replicates) in predicting future performance. The National Turfgrass Evaluation Program (NTEP) relies on the cell means model (i. e., full model or AMMI-F) for estimates to predict turfgrass quality performance. The objective of this study is to compare under field conditions AMMI winners (AMMI adjusted means) with standard NTEP winners (AMMI-F) from the 1995 Kentucky bluegrass (Medium-High Maintenance) variety trial. Model validation studies indicated that the AMMI-5 model (five interaction axes) was the most predictively accurate model with a statistical efficiency relative to AMMI-F of 1.25. Validation studies were based on analysis of turf quality (TQ) ratings for 103 genotypes averaged over year (1996 to 1999). Five winning genotypes suggested by each model (AMMI-5 and AMMI-F) were planted in the fall of 2003 at nine NTEP locations in the USA and Canada. This is the first year of a three-year field evaluation. Only six of the nine test locations reported TQ data (1 to 9 scale, 9=highest) for year 2004. AMMI-F and AMMI-5 predictions were uncorrelated with actual yearly TQ at four of the six test locations. AMMI-5 adjusted means and actual yearly TQ were significantly and positively correlated at two of the six test locations. Alternatively, AMMI-F predictions were significantly but negatively correlated with actual yearly TQ at two test locations. These results indicate that AMMI-F predicted winners were actual losing genotypes and in turn AMMI-F predicted losing genotypes were actual winning genotypes. AMMI-5 model predictions of actual turf quality in 2004 were more accurate than AMMI-F (NTEP model).