Sugar content (SC) is one of the important internal qualities of Nanshui pears. In this research, Visible/near infrared (Vis/NIR) spectroscopy was used to detect SC of Nanshui pears on-line. Transmission speed of Nanshui pears was 0.3 m/s, and USB4000 spectrometer was used to acquire the spectra of Nanshui pear samples in the wavelength range of 470～1 150 nm. Then three variable selection methods were used to select sensitive wavelength variables, and partial least squares (PLS) was used to develop calibration models of SC for Nanshui pears, also performance of calibration models was compared. The results indicate that Vis/NIR spectroscopy combined with variable selection method is feasible for on-line detection of SC for Nanshui pears. Competitive adaptive reweighted sampling (CARS) method is superior to uninformative variable elimination (UVE) and successive projections algorithm (SPA) methods. CARS method can simplify calibration model and improve performance of calibration model. The correlation coefficients in prediction and root mean square errors of prediction (RMSEPs) of full-PLS and CARS—PLS models of SC for Nanshui pears are 0.940,0.951 and 0.467%,0.420%, respectively.