Traditional finite-control-set model predictive control is widely applied to complex nonlinear systems such as modular multilevel converters (MMC) due to its capability for multi-objective control. However, its performance deteriorates under parameter mismatch and sensor noise conditions. To address these issues, this paper proposes a model-free predictive control strategy for MMC based on a Kalman filtering-quasi-resonant extended state observer (KF-QRESO) to enhance system robustness against parameter mismatch and sampling disturbances. First, the discrete mathematical model of MMC under parameter mismatch is analyzed, and a composite KF-QRESO observer is constructed. The Kalman filter (KF) is used to suppress sampling noise, while the QRESO accurately estimates periodic AC state variables and compensates them within the KF state equations. Then, the composite observer reduces the parameter dependence of the control system, achieving precise estimation. The tracking ability and stability of the observer for periodic signals are also studied. Next, the composite observer is integrated with model-free predictive control to improve performance under parameter mismatch and sampling noise conditions. Finally, MATLAB/Simulink simulations and prototype experiments validate the method’s effectiveness and correctness.