Near-infrared (NIR) spectroscopy combined with multivariate calibration methods is extensively applied in quantitative analysis of complex samples due to its advantages of rapidity and non-destructiveness. However, peak overlapping and variable redundancy in NIR spectra reduce the prediction accuracy and generalization ability of the model. Thus, it is crucial to select variables from the whole spectra before modeling. In this research, as a first attempt, tyrannosaurus optimization algorithm (TROA) was discretized and introduced for spectral variable selection and combined with the partial least squares (PLS). The method was validated by four NIR spectral datasets of tablets, oil blends, orange juice and soil samples. The key parameters, including the number of latent variables (LVs) for PLS, the transfer function and the number of iterations of TROA, were optimized, respectively. Compared with full-spectrum PLS, uninformative variable elimination-PLS (UVE-PLS), Monte Carlo-UVE-PLS (MC-UVE-PLS), randomization test-PLS (RT-PLS), competitive adaptive reweighted sampling-PLS (CARS-PLS), grey wolf optimizer-PLS (GWO-PLS), and whale optimization algorithm-PLS (WOA-PLS), TROA-PLS demonstrated notable advantages in both model simplicity and prediction accuracy for the four datasets. Therefore, TROA is an efficient and reliable method for NIR spectral variable selection, providing a new method for spectral modeling of complex samples.