Abstract: The moldy deterioration of grain severely impacts its quality and food safety, while conventional detection methods suffer from slow speeds and requirements for extensive professional laboratory equipment and operators, among other drawbacks. Near-infrared spectroscopy (NIRS) analysis technology, characterized by rapid analysis, non-destructiveness, good test reproducibility, ease of online analysis implementation, and simplicity of operation, represents a promising method for rapid detection. This study established a mold detection model based on NIRS analysis technology to conduct rapid identification and prediction research on the near-infrared spectra of rice with varying degrees of moldiness. The aim was to develop a qualitative and quantitative model that can swiftly differentiate moldy rice. The study involved 960 samples of rice with different degrees of moldiness from four varieties (Mudanjiang 27 from 2018, Mudanjiang 27 from 2019, Longjingchanglixiang, and Muxiang No. 1). These were used to investigate qualitative discrimination models for varying degrees of moldiness. Among them, the first-order derivative combined with 9-point smoothing and factorization resulted in a qualitative discrimination model with high accuracy, with a mean S-value greater than 1, indicating excellent discrimination. The accuracy of the model was validated through leave-one-out cross-validation, achieving 93% accuracy. Additionally, 300 samples of rice with four different degrees of moldiness were used for quantitative model research on total mold colony counts based on NIRS. The quantitative discrimination model was established through vector normalization and partial least squares (PLS) quantitative analysis. The cross-validated root mean square error (RMSECV), coefficient of determination (R²), ratio of performance to deviation (RPD), and root mean squared error of prediction (RMSEP) were 0.47, 90.45, 3.24, and 0.45, respectively. The first three parameters demonstrate high model precision, while the latter indicates high model accuracy. Therefore, this model achieves a well-precision rapid prediction model for moldiness. After optimizing the analytical methods, the established mold discrimination model revealed that "moldiness" is the dominant factor affecting near-infrared spectral changes, while rice variety and year have relatively minor influences.