We developed a Fast Acquisition and Reconstruction CEST (FAR-CEST) method at 3T human scanners, based on a deep learning approach. A 10X accelerated acquisition was achieved, which under-sampled K-space using a randomized Cartesian pattern of variable density. To fully utilize the correlation among saturation offset dimension, especially to compensate for sparsely-sampled K-space edge, a 3D-Res-Unet model was trained for reconstruction. Results on healthy adult brain suggested that FAR-CEST can produce high quality saturation-weighted images and Z-spectra，but the CEST contrast slightly altered. The highly-acceleration feature of FAR-CEST has been initially validated, yet still require improvement on reconstruction accuracy.