This paper introduces a non-local tensor sparse coding approach for multi-image super-resolution in magnetic resonance imaging. This approach is composed of four steps: (i) non-local clustering of the similar subtensors, (ii) tensor sparse dictionary learning, (iii) tensor sparse coding and (iv) subtensor regularization. Using the Tucker decomposition, the image reconstruction problem is transformed into learning of sparse dictionaries along the three modes and core tensor sparse coding for each cluster, viewed as tensor. With the proposed approach, reconstruction is achieved with only two low-resolution images, which is a major advantage compared to other multi-frame reconstruction techniques. Flexible conditions for exact recovery are provided. This is also a major advantage of the proposed approach, that will facilitate the clinical implementation of our algorithm, allowing physicians to obtain images very quickly after acquisition, with both practical (convergence analysis) and theoretical guarantees (theorems). The experiments on a set of real quality test phantom and brain datasets show the competitive performance of the proposed approach with a significant gain of time compared to other state-of-the-art methods.