Serial image acquisition is always required for quantitative and dynamic MRI, in which the spatial and temporal resolution as well as the contrast is restricted by acquisition speed. Herein we employed a novel transformer based deep-learning method for reconstruction of highly under-sampled MR serial images. The network architecture and input were designed to take advantages of Transformer’s capacity on global information learning along the redundant series dimension. SMR-transformer was compared with 3D Res-U-Net, on a self-acquired CEST dataset, a cardiac dataset from OCMR and a diffusion dataset from HCP. Both images and quantitative comparison indicated SMR-Transformer allowed fast serial MRI.