Next point-of-interest (POI) recommender sys-tems that recommend the user's next POI according to past check-ins history are widely deployed on mobile devices. Recently, there has been remarkable progress on sequential models which improve recommendation performance by learning in-sequence check-in interactions, yet such models are memory and computationally costly. To fill the gap, we aim to develop an effective and efficient next POI recommen-dation algorithm on sequential models. For effectiveness, the spatial and temporal check-ins information are fed into the sequential model through our proposed self-attention structure, which fully utilizes the relative distance information to calculate the check-ins relationship and compensates by the check-in timings. For efficiency, we select representative elements in the self-attention structure to roughly represent the whole self-attention score, thereby reducing the memory and computational time costs. Our extensive experiments on real-world datasets show that our proposed method has state-of-the-art performance among all baselines, and the memory and time complexities of the self-attention component re-d uced from to .