Nondestructive, rapid, and accurate detection of nutritional compositions in sorghum is crucial for agricultural and food industries. In our study, the crude protein, tannin, and crude fat contents of sorghum variety samples were taken as the research object. The visible near-infrared (VIS-NIR) hyperspectral of sorghum were measured by the indoor mobile scanning platform. The nutritional components were determined using chemical methods to analyze the differences in nutritional composition among different varieties. After preprocessing the original spectral, the competitive adaptive reweighted sampling (CARS) and bootstrapping soft shrinkage (BOSS) algorithms were used to coarsely extract the key variables. Subsequently, the iteratively retains informative variables (IRIV) was employed to assess the importance of these key variables, resulting in explanatory wavelength sets for crude protein, tannin, and crude fat. Finally, the partial least squares (PLS), back propagation (BP) and extreme learning machine (ELM) were utilized to establish detection models. The results indicated that the optimal wavelength variable sets for crude protein, tannin, and crude fat contained 41, 38, and 22 wavelength variables, respectively. The CARS-IRIV-PLS, BOSS-IRIV-PLS and BOSS-IRIV-ELM were suitable for detecting crude protein, tannin and crude fat, respectively. Meanwhile, the Rp2, RMSEp and RPDp values of the model were 0.69, 0.80% and 1.80, 0.88, 0.22% and 2.84, 0.61, 0.32% and 1.61, respectively. These detection models can be used for the effective estimation of the nutritional compositions in sorghum with VIS-NIR spectral data, and can provide an important basis for the application of food nutrition assessment.