An optimized partial least squares (PLS) quantitative prediction model is designed for the nondestructive determination of apple acidity by visible and near infrared spectroscopy (Vis-NIRS). Firstly, Savitzky-Golay smoothing combined with wavelet transform is used to preprocess the spectral data. Then the successive projections algorithm (SPA) is used to generate a modeling set, and a modeling candidate set is also generated at the same time by the competitive adaptive reweighted sampling (CARS) and SPA. Furthermore, the wavelength variable is successively selected from the modeling candidate set to the modeling set, and a prediction model is established finally according to the modeling set until the change of the determination coefficient tendsto be stable, thus achieving a best-fit model. The experimental results show that when apple acidity is predicted, the determination coefficient and the relative percent deviation of the optimized PLS model reaches 0.9776 and 6.6812 respectively, and thenumber of selected wavelength variables is reduced from 129 to 36, which is obviously superior to that of SPA and CARS. The designed model not only ensures the model accuracy, but also reduces its complexity, which provides an important reference for theestablishment of online nondestructive determination model of apple acidity.