The use of near infrared spectroscopy (NIRS) as an alternative to the techniques commonly employed in the study of forage composition needs to be explored. The objective was to construct calibration curves to predict the physical-chemical composition of tropical grasses (Brachiaria brizantha (Hochst. ex A. Rich.) Stapf 'Marandu', 'Piata'; B. decumbens Stapf, Panicum maximum Jacq. 'Coloniao'), by NIRS and compare two multivariate regression method. Forage samples were analyzed for crude protein (CP), acid detergent fiber (ADF), neutral detergent fiber (NDF), ash, ether extract (EE), lignin, and moisture. The values obtained by the Official Methods of Analysis of Association of Official Agricultural Chemists (AOAC) were reference values for the creation of multivariate calibration models. The samples were scanned on the NIRS. The multivariate calibration models were created by the partial least squares (PLS) method and by the multiple linear regression (MLR) method. The predictive capacity of the models was evaluated by the correlation coefficient (R) and parameters of the mean squared deviation (RMSE). When the MLR was used, only the prediction model of ash (R = 0.82) of the P. maximum, EE (R = 0.87) and moisture (R = 0.90) of ' Piata' showed approximate predictive capacity, for the other components R indicated good prediction. After the validation of the models developed by the PLS regression method, the CP (0.78-0.91), NDF (0.88-0.95), lignin (0.85-0.91), and moisture (0.79-0.96) predictions presented good results. The NIRS technique can be used to determine the physical-chemical composition of tropical grasses. The MLR multivariate regression method as well as PLS can be used to predict the physical-chemical composition of tropical grasses. Key words: Brachiaria, C4 forage, forage composition, multiple linear regression, multivariate statistics, near infrared spectroscopy.