To reduce the torque ripples of permanent magnet synchronous machine (PMSM) drive systems with model predictive control (MPC), an improved MPC is proposed based on torque ripple minimization. The optimal active vector is selected from the control set for current control. Considering voltage vectors and torque responses in one control period, a cost function of torque ripples is designed to calculate the optimal duration of the active vector. The selected active vector and a zero vector are combined according to the optimal duration to determine switching control signals of the inverter. Simulation results indicate that the proposed method reduces torque ripples and has high dynamic responses of torque at different operation conditions.