To address issues of low accuracy and low robustness in predicting heterogeneous samples, this study focuses on Polygonatum sibiricum polysaccharide and proposes a model transfer algorithm based on homogeneous samples. By incorporating a hybrid modeling strategy, different physical-state mixed prediction models were established. Stacking ensemble learning was employed to establish base prediction models, and a radial basis function (RBF) neural network was introduced as the transfer function in transfer near-infrared spectroscopy. It was used to fit the nonlinear mapping relationship of spectra from samples with different physical states. By adjusting the size of absorbance matrix window, the network fitting effect was optimized and the near-infrared spectroscopy transfer function was determined. Results indicate that the mixed prediction model corrected using the RBF achieves fitting coefficient (R²) of 0.991, root mean square error (RMSE) of 0.497%, and mean absolute error (MAE) of 0.383% for testing set. Proposed nonlinear transfer algorithm effectively manages sample complexity, reduces the effects of sample surface morphology and moisture on modeling, and enhances the accuracy and generalizability of mixed prediction model for Polygonatum sibiricum polysaccharide content.