The Extended Min-Sum (EMS) algorithm is one of the most commonly used non-binary low-density paritycheck (NB-LDPC) decoding algorithms, with excellent decoding performance and moderate complexity. In this letter, we propose an EMS algorithm based on model-driven deep learning, called neural EMS (NEMS) algorithm. The NEMS algorithm can be regarded as an extension of the EMS algorithm by adding trainable weights to variable nodes and check nodes. When the added learning weights are all 1s, the NEMS algorithm degenerates into an EMS algorithm. The experimental results show that the BER performance of the NEMS algorithm outperforms that of the EMS algorithm.