<p>With the rapid development of the Artificial Intelligence of Things (AIoT), mobile edge computing (MEC) has become an essential technology underpinning AIoT applications. However, multi-angle resource constraints, multi-user task competition, and the complexity of task offloading in dynamic MEC environments pose new technical challenges. To address these, we propose a user-centric deep reinforcement learning (DRL) model splitting (UCMS) inference scheme. This scheme combines a user-server co-selection algorithm with a UCMS_MADDPG-based offloading algorithm to realize efficient inference responses in dynamic environments with multi-angle resource constraints. Specifically, we formulate a joint optimization model that integrates resource allocation, server selection, and task offloading, aiming to minimize the weighted sum of task delay and energy consumption. After decoupling the optimization, the user-server association is handled through a co-selection algorithm. To address the mixed decision problem, we design an algorithm centered on user pre-decision that splits the action space into user-side and server-side components to coordinate continuous and discrete decision outputs. In addition, a priority sampling mechanism based on a reward-error trade-off is introduced to enhance experience replay. Simulation results show that the proposed UCMS_MADDPG-based offloading algorithm demonstrates superior overall performance compared with other benchmark algorithms in dynamic environments.</p>