A new approach is established to determine the contents of rifampicin (RMP), isoniazide (INH) and pyrazinamide (PZA) in five varieties of anti-tuberculosis tablets,i.e. rifampicin tablets, isoniazide tablets, pyrazinamide tablets, rifampicin and isoniazide tablets and rifampicin isoniazide and pyrazinamide tablets. Near infrared spectroscopy (NIRS) with radial basis function neural network (RBFNN) is applied to establishing the analytical protocol for the determination of the contents of RMP, INH and PZA in the five anti-tuberculosis tablets. The root mean square error of calibration set obtained by cross-validation (RMSECV) is used as the evaluation of the model. The optimal quantitative analysis models are resulted from selecting the most effective spectral region,suitable topological parameter and spread constants. RMSECV of the optimum models are 0.0127, 0.0104 and 0.0078. Using these optimum models for determination of the RMP, INH and PZA contents in prediction set, the root mean square error of prediction set (RMSEP) are 0.0125, 0.0109 and 0.0103. Internal cross-validation and external validation are verified and the method has a high accuracy and can meet the accuracy in simultaneous determination of RMP, INH and PZA contents in 5 varieties of anti-tuberculosis tablets.