To develop a universal blueberry soluble solids content (SSC) detection model adaptable to randomly arranged production line scenarios, this study collected the SSC values of 300 blueberry samples and their hyperspectral data on four surfaces. Partial least-squares regression, support vector regression, and Long Short-Term Memory (LSTM) were applied to construct both local and universal models based on multi-surface spectral data. The results indicated that the local models constructed using singlesurface spectral data performed poorly in cross-surface prediction. In contrast, the universal models, constructed based on multisurface spectral data, significantly improve generalizability. Specifically, prediction correlation coefficients of the universal LSTM model on the external validation set for the four surfaces were 0.963, 0.954, 0.964, and 0.962, respectively. The root-mean-square errors of the predictions were 0.513%, 0.521%, 0.472%, and 0.483%, respectively, and the residual prediction biases were 3.32, 3.27, 3.61, and 3.52, respectively. This study validates the feasibility of constructing a universal SSC detection model based on multi-surface spectral data and demonstrates the potential application of the LSTM model in blueberry SSC detection. The LSTM universal detection model constructed in this study provides a novel technical solution for the non-destructive detection of blueberry quality on the production line. The multi-surface spectral fusion method can provide a theoretical reference for the development of quality detection models for other small-sized fruits.