Tight sandstone reservoirs are characterized by complex pore structures, strong heterogeneity, low porosity, low permeability, and highly variable permeability. Accurate well-log-based permeability prediction is therefore critical for their efficient development. Nuclear magnetic resonance (NMR) logging, which provides porosity information independent of rock composition, serves as a valuable tool for this purpose. However, conventional NMR-derived models such as SDR(Schlumberger-Doll research) and Coates have inherent limitations. Therefore, based on the key parameters of the Coates model, using four parameter wells in the Wuguo-Jixian exploration area of the Ordos Basin as the data set, a deep neural network model (DNN) and a long short-term memory network model(LSTM) were constructed for permeability prediction using nuclear magnetic resonance logging. Experimental results demonstrate that the DNN-NMR model achieves good performance in static parameter fitting (R² = 0.872), representing a 14% accuracy improvement over the Coates model. The LSTM-NMR model performs best, achieving the highest accuracy (R² = 0.916). The research results provide a high-precision and robust solution for the quantitative NMR-log-based prediction of permeability in tight sandstone reservoirs.