Servo drives require high dynamics and reliability on speed control. Conventional cascade linear controllers suffer from the proportional-integral parameters tuning work and low dynamic response, due to their cascaded structure. In this paper, an improved model predictive direct speed control is proposed with rapid speed tracking and very small speed offset. The new control scheme eliminates the cascaded structure by predicting the future speed in discrete steps. The optimal voltage vector to control the motor is then selected according to an evaluation criterion for speed and flux tracking. To reduce the system cost and improve the reliability, a load torque observer is adopted to estimate the actual load torque. Besides, to avoid torque oscillations and overshoots during rapid speed variation, a torque suppression factor is incorporated into the cost function. Furthermore, a myopic prediction correction method is developed to enhance both the dynamic and steady-state responses. Simulation and hardware-in-the-loop results are presented to validate the effectiveness of the proposed method.