The chestnut (Castanea) is an important fruit in Europe and Asia. As a highly variable fruit, its quality is graded according to nutrition components, especially according to the sugar content, which are traditionally measured by using chemical methods. However, the traditional methods are time-consuming, laborious, and expensive. Here, we analyzed the sugar content of intact and peeled chestnuts by near-infrared spectroscopy. The spectra of intact and peeled chestnut samples were collected in the wavelength range from 833 nm to 2500 nm. The Sample Set Partitioning based on joint X–Y distances was used when the calibration and validation subsets were partitioned. The predictive models for intact and peeled chestnut samples respectively, were developed using partial least squares (PLS) regression based on the original spectra and the spectra derived from different pretreatments. The PLS models developed from the spectra of peeled samples gave accurate predictions. The correlation coefficient (R2) of the optimized model for calibration set and validation set were 0.90 and 0.86. Although the models established on the spectra of intact samples did not perform excellently, they were still qualified to measure sugar content of the chestnut kernel. The correlation coefficient (R2) of optimized model for calibration set and validation set were 0.89 and 0.59. These results suggested that NIR spectroscopy could be used as a fast and accurate alternative method for the nondestructive evaluation of sugar content in chestnuts during orchard and post-harvest processes.