Recent studies have demonstrated that increasing the size of the self-attention window for Transformer-based image super-resolution models can remarkably enhance model performance. However, the quadratic computational complexity of self-attention is unacceptable, particularly as the self-attention window increases. To address this problem, linear attention provides a more efficient alternative. Nevertheless, its performance is not as satisfactory compared to Softmax Attention. In this paper, we propose an efficient image super-resolution network(MASR) that combines Linear Attention and Softmax Attention. Our approach significantly reduces resource consumption while preserving or even enhancing image quality. On the whole, MASR consists of four parts: shallow feature extraction module(SFE), CNN and Linear Attention Transformer hybrid blocks (CLATs), CNN and Softmax Attention Transformer hybrid blocks(CSATs), and a high-quality image reconstruction module(IR), respectively. Extensive experiments demonstrate that our MASR outperforms existing methods.