The size difference of the fruit itself results in poor robustness and low accuracy of the model for predicting the content of sugar. Eliminating the influence of fruit size differences is of great significance to improve the accuracy of fruit sorting models. The NIR spectra of apples with different fruit diameters were collected by an online NIR spectroscopy detection device， and the partial least squares （PLS） models of the content of sugar in apple （SSC） were established after various pre-processing of the spectra. The modeling sets in the diameter of apple fruit group with 75-85 mm were used to predict the prediction set samples in the diameter of apple fruit group with 65-75 mm and 85-95 mm， respectively. The modeling and prediction sets in the diameter of apple fruit group with 65-75 mm， 75-85 mm， and 85-95 mm were used as the modeling and prediction sets of the mixed size model for predicting the content of sugar in apple， respectively. The model was simplified by using the feature spectral selection algorithm. The correlation coefficient R_p was increased from 0.805 to 0.943， the root mean square error value RMSEP was reduced from 0.778 to 0.480， and the RPD was increased from 0.96 to 3.05 compared with the optimal set of model for predicting the content of sugar in apple established when the modeling set and the prediction set had different fruit diameters. It is indicated that simplifying the general model established can reduce the effects of apple size on the model for predicting the content of sugar in apple and improve the prediction performance of the model.