Based on the hyperspectral imaging (HSI) technique, this paper attempts to test the saccharinity of three varieties of cherry tomatoes in a nondestructive manner. The cherry tomato samples of the three varieties were collected, and kept at room temperature for 12h. Then, the spectral curves of the samples were obtained between the wavelengths of 914.91nm and 1,661.91nm. After that, the feature bands were extracted by three algorithms, namely, competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA) and SPA-CARS. The samples were divided into a correction dataset and a prediction dataset at the ratio of 2:1. Next, the feature bands extracted by the three algorithms were combined with the partial least squares (PLS) and least squares-support vector machine (LS-SVM) into six saccharinity prediction models. Finally, the prediction results of the six models were compared, revealing that the CARS-LS-SVM achieved the best performance with a prediction accuracy of >92%. The evaluation indices of this model are as follows: the correlation coefficient of correction dataset (R), 0.9696; the correlation coefficient of prediction dataset (R), 0.9220; the root mean square error of correction dataset (RMSEC), 0.2768; the root mean square error of prediction dataset (RMSEP), 0.4390. The research results lay the basis for industrial grading of saccharinity of cherry tomatoes in a nondestructive manner.