Deadbeat Predictive Current Control (DPCC) has become a popular choice in permanent magnet synchronous machine (PMSM) drive systems thanks to its outstanding dynamic and steady-state performance. However, parameter mismatch will lead to current tracking errors and large torque fluctuation. To enhance the parameter robustness of PMSM drive systems, an improved model-free DPCC (IMFDPCC) is proposed in this article. Based on ultralocal model, the proposed method requires only system input and output instead of any motor parameters. Moreover, the proposed parameter-adaptive ultralocal model can adaptively adjust the coefficient of input and the values of disturbance by recursive least squares (RLS) algorithm and extended state observer (ESO), respectively. The combination of RLS and ESO simplifies the calculation and tuning work on one hand, and makes the motor drive system perform well in different situations on the other hand. Compared to conventional DPCC and model-free DPCC, the proposed method achieves better performance in current tracking errors and harmonics.