This paper presents an improved model predictive control strategy to reduce torque ripples in conventional direct torque control (DTC) of a permanent magnet synchronous motor. Nowadays, DTC has been widely used in industrial applications thanks to its simple structure and high dynamic performance. However, hysteresis controllers cause high torque ripple and deteriorate control performance. In this paper, hysteresis controllers and traditional DTC switching table are replaced by a model predictive controller to achieve an online optimization for the required driving voltage space vector. Furthermore, a pre-selective process is designed to filter out the inappropriate voltage vectors and reduce computational burden caused by the eight basic voltage space vectors. Finally, an optimal duty ratio modulation method is proposed to determine appropriate zero voltage vector and the duty ratio of the selected voltage vector in each control period for torque ripple reduction. By implementing the proposed DTC, torque ripple is reduced and the approximated switching frequency can be minimized to retain the advantages of a traditional DTC; meanwhile, the torque response is still quick. Both simulation and experimental results exhibit the effectiveness of this proposed approach.