Computational burden is a major hurdle for practical implementation of finite-state predictive torque control (FS-PTC) of motor drive fed by a multilevel inverter. One of the reasons of computational complexity is that all voltage vectors are evaluated for prediction and actuation. This paper proposes a reduced number of voltage vectors for the prediction and actuation, which are called prediction vectors in FS-PTC. The performance is investigated for a three-level neutral-point clamped inverter fed motor drive in terms of torque and flux response, stator current total harmonic distortion, robustness, average switching frequency, and neutral-point voltage variation. The number of prediction vectors is reduced based on the position of stator flux and the deviation in stator flux from its reference. Experimental results confirm that the computational burden could be reduced by 38%, while the dynamic performance is comparable with the conventional all voltage vectors based FS-PTC.