Soluble solid content (SSC) is a critical indicator of ‘Red Fuji’ apple quality, directly governing fruit grading and maturity assessment processes. Conventional SSC measurement by refractometry is destructive and time-consuming, rendering near-infrared diffuse reflectance spectroscopy (NIR-DRS) a promising nondestructive alternative. In this study, a low-cost and compact embedded spectrometer named as DLP NIR-scan Nano EVM was used to acquire NIR-DRS spectra of ‘Red Fuji’ apples for SSC prediction. To improve prediction accuracy, we combined spectral preprocessing with machine learning methods. The dataset was cleaned using Monte Carlo outlier detection, and samples were divided into calibration and validation sets via Kennard–Stone (KS) and joint X-Y distance (SPXY) algorithms. Among preprocessing methods tested, a 12-point second derivative performed best when paired with KS partitioning. For feature-wavelength selection on the preprocessed KS data, competitive adaptive reweighted sampling, Monte Carlo uninformative variable elimination, and Random Frog were applied to the second-derivative spectra. Partial least squares regression (PLSR) models were then built using both full-spectrum data and four sets of selected wavelengths. The best preprocessed PLSR model achieved R2c = 0.916, RMSEC = 0.4093%, R2p = 0.8632, and RMSEP = 0.537%. These results demonstrate that NIR-DRS, combined with appropriate preprocessing and modeling strategies, offers a reliable, rapid, and nondestructive method for apple SSC quantification, paving the way for portable, cost-effective instruments for commercial fruit quality monitoring.