Permanent magnet synchronous motor(PMSM) drives are attractive for sustainable energy conversion due to their high efficiency and power density. The model predictive direct torque control(MPTC) of PMSM drives supports desirable control requirements including high dynamic performance, low torque/flux ripples and zero overshoots of torque. However, MPTC’s shortcomings lie in its sensitivity to both parameter uncertainties and measurement noises. Therefore, this paper proposes overcoming these challenges by a novel model-free predictive direct torque control using hybrid parallel-cascade extended state observer. The proposed method is validated by theoretical frequency-domain analysis and time-domain experiments. The real-time hardware-in-the-loop test results validate the superior measurement noise attenuation, and parameter robustness of the proposed method over both the classical MPTC and the conventional ESO-based model-free predictive torque control.