The recently proposed smooth twin support vector regression, denoted by STSVR, gains better training speed compared with twin support vector regression (TSVR). In the STSVR, sigmoid function is used for the smooth function, however, its approximation precision is relatively low, leading to the generalization performance of STSVR is not good enough. Moreover, STSVR has at least three parameters that need regulating, which affects its practical applications. In this paper, we increase the regression performance of STSVR from two aspects. First, by introducing Chen-Harker-Kanzow-Smale (CHKS) function, a new smooth version for TSVR, termed as smooth CHKS twin support vector regression (SCTSVR) is proposed. Second, a binary particle swarm optimization (PSO)-based model selection for SCTSVR is suggested. Computational results on one synthetic as well as several benchmark datasets confirm the great improvements on the training process of proposed algorithm.