When using near-infrared spectroscopy for detection, the spectral band contains significant noise and scattering, which affect the stability of the model. Based on the competitive adaptive reweighting (CARS) and mutual information (MI) algorithms, a partial least-squares (PLS) regression model was established to detect the soluble solid content (SSC) in apples. The diffuse reflectance spectrum data of 120 samples at 800?2400 nm were obtained using a spectrometer. After preprocessing, 96 samples were randomly selected as the calibration set for modeling, and 24 samples were selected as the prediction set for prediction using the Kennard-Stone (KS) algorithm. Next, the full-band PLS model, CARS-PLS model, and MI-PLS model were established for comparative analysis. The results show that the coefficient of determination R² of the PLS model is 0.8511, the root-mean-square error of calibration (RMSEC) of the model and the root-mean-square error of prediction (RMSEP) are 0.9413 and 1.1915, respectively. The number of characteristic wavelength point variables screened by the CARS algorithm reduces from 303 to 12, a decrease of 96.03%. The coefficient of determination R² of the PLS model is 0.8746, an increase of 2.76%. The RMSEC and RMSEP values are 0.864 and 0.9757, respectively. The MI-PLS model contains 56 characteristic wavelength points, and the selected wavelength accounts for 18.49% of the total wavelength. R², RMSEC, and RMSEP are 0.9218, 0.6822, and 0.8235, respectively. Compared with CARS-PLS, the number of characteristic wavelengths of MI-PLS increases by 64.55%, and the coefficient of determination R² increases by 0.0472. Therefore, the CARS and MI algorithms can overcome the problems of noise and scattering of spectral data and can be effectively used for screening characteristic bands. The established model is suitable for determining the SSC in apples.