Near infrared (NIR) spectral analysis with a NIRSystems 6500 monochromator was applied to evaluate accuracy of predictive models for forage quality in clone breeding processes of the original Bulgarian varieties over different cuts and years. The varieties were the perennial grasses: Dactylis glomerata L., Festuca arundinacea Schreb. and Bromus inermis Leyss. Global calibrations for the 418 perennial grass samples and specific calibrations for each single grass species and internal cross-validations were performed by the PLS regression method. The effect of different spectral data pre-treatments was investigated on the residual standard errors of the NIR predictive models. Among 60 calibration equations, the model with the lowest SECV value was retained for each parameter in each database. No particular data pre-treatment was really better than the other ones. Generally, the best results of the global calibrations were obtained with SNVD and MSC. For the specific calibrations, SNVD and WMSC were the best treatments. In both cases, the first or second derivatives were needed after the first pre-treatment. Chemical composition and in vitro enzymatic digestibility of clones were predicted with accuracy similar to that of classical laboratory methods. For the cell wall component contents, the standard errors of cross-validation SECV(%DM) ranged from 0.49 for ADL ( Festuca) to 2.02 for NDF ( Dactylis). The digestibilities of dry and organic matter, IVDMD and IVOMD, were estimated with SECVs from 2.6 to 3.0%, the relative intake, from 0.06 to 0.09 rel% body weight and the relative feeding value, from 4.39 to 5.64 rel%. The global calibration models offer an acceptable accuracy for the estimation of the cell wall nutrient contents, the digestibility and the nutritive value. The standard errors of prediction of specific single species calibrations with smaller numbers of terms were lower in 60% of the cases than those obtained from the best global calibrations with higher numbers of terms. On average, SECVs from specific calibrations are better than those from global calibrations, but the differences are quite small, and for the prediction of totally new samples (new crops, another year), the global calibrations will detect less outlier samples. Even with very high variability between cuts and years, NIR spectroscopy is able via ANOVA GL Models to sort clones on their feeding value and to provide relevant information for the breeding programmes.