Near-infrared spectroscopy was used to quantitatively analyze the fat content of walnut beverage. At the same time, modeling variables were optimized and modeling methods were compared to optimize the best model. In order to eliminate the influence of scattering on the spectrum, the data are preprocessed by the standard normal transformation (SNV) method. The preferred characteristic wavelengths of genetic algorithms (GA) combined with backward interval partial least squares (BiPLS) were used as input variables of partial least squares (PLS) and least squares support vector machine (LS-SVM) respectively to establish model of fat content in walnut beverage. The R2, RMSEP and RPD were used to evaluate the effect of spectral band selection method on the construction of fat index model in walnut beverage and determine the best modeling method. The results showed that the variable selection could optimize the model. 150 and 30 variable points corresponding to the characteristic absorption peaks of the fat components in walnut beverage samples were selected by BiPLS and GA-BiPLS methods, respectively, accounting for 10% and 2% of the full spectrum. The RMSEP value of the PLS model decreased from 0.049 to 0.043 and 0.040, respectively, and the R2 increased from 0.964 to 0.973 and 0.974. The range error ratio RPD increased from 4.88 to 5.62 and 6.00, and the principal component number also decreased to varying degrees. The method of variable selection could reduce model dimensions and improve model accuracy. Compared with the PLS model, the R2, RMSEP and RPD values of the LS-SVM model showed better results, reaching 0.986, 0.036 and 6.52, respectively. The LS-SVM model has higher accuracy and stability than the PLS model. Since PLS is a classic linear modeling method, the nonlinear factors in the sample data set are ignored in the process of building the model. However, there was a complex nonlinear relationship between fat content and near-infrared spectral information, which is due to the interference of noise, background and other factors in the spectral measurement process of walnut beverage samples and the interaction between various indicators. The LS-SVM method could enhance the correlation between spectral variables and index concentration, so that the established model has better accuracy and universality. It shows that in the actual production, the LS-SVM method has excellent feasibility, which reflects its great potential in the analysis of the quality of walnut beverage. Based on the LS-SVM method, the quantitative analysis model of walnut fat content has accurate and stable characteristics, which can provide technical reference for the quality monitoring of walnut beverage production, and provide a new idea for the analysis of beverage quality.