In order to distinguish tea grade by using visible-near-infrared spectroscopy technique, Meitan Cuiya tea was used as materials in this study. The spectral data of all different grades Cuiya samples were collected. Firstly, Savitzky-Golay smoothing(SG), multiple scattering correction(MSC), standard normal variable transformation (SNV), first derivative, second derivative, detrending and other pretreatment methods were used to process the original spectral data of the samples. Then, the partial least squares regression (PLSR) model was established based on different preprocessing methods and raw data. The influence of different pretreatment methods on the modeling model was also studied. The results showed that the modeling of SG smoothing pretreatment method had the best effect. In order to simplify the model, three characteristic wavelength selection methods, the stepwise regression analysis (SWR), successive projection algorithm (SPA), and competitive adaptive re-weighting (CARS) were used to select the characteristic wavelength, which would be the pretreatment before the SG smoothing. Finally, PLSR modeling was performed based on the characteristic wavelengths selected by different feature wavelength algorithms. The results showed that the model based on the CARS method had the best prediction effect, with the correlation coefficient of 0.9739 and the calibration standard deviation of 0.2250. The model greatly reduced the number of independent variables, simplified the previous model, and achieved a good prediction effect, which provided a new, quick and effective method for the classification of Cuiya grades.