Modular multilevel converters (MMCs) are mostly used in active power filters (APFs) due to their high modularity and easy scalability. Furthermore, model predictive control (MPC) has become one of the preferred methods in many modern control applications due to its good control effect and strong robustness. Aiming at the problem that the computational load of the traditional MPC rises significantly with the increase in the number of submodules (SMs), this paper proposes an improved backward MPC strategy for modular multilevel active power filter (MMC-APF). This strategy is based on nearest level modulation (NLM) to reversely predict the number of SMs pre-inserted in the upper and lower bridge arms, then expands the number of SMs inserted in each phase from N to [N − 1, N + 1], and the optimal number of SMs finally inserted in the upper and lower bridge arms is determined by minimizing the circulating current. A sorting algorithm is used to obtain the switching states of each SM, while maintaining the balance of capacitor voltages and reducing the switching frequency. The proposed strategy does not need to design weighting factors and greatly reduces the computational load of MPC. Simulation results show that the control strategy proposed in this paper has a small computational load and has good performance in harmonic current compensation, verifying the effectiveness of the proposed control strategy.