<p>Distributed edge computing systems require real-time execution of computationally intensive deep neural network inference for a large number of concurrent tasks. Meeting stringent latency constraints while controlling operational cost poses significant challenges to existing edge intelligence solutions. Performing complete model computation locally is often constrained by the limited computational capacity of end devices, making it difficult to meet the accuracy requirements of complex models. Conversely, fully offloading computation to edge servers is frequently hindered by prohibitive energy consumption, undermining economic efficiency. To address these issues, this paper proposes a Reinforcement Learning-based Adaptive Model Partitioning approach (RLAMP), which first selects suitable models based on real-time computational requirements and device latency constraints to ensure minimal computational capability. It then dynamically adjusts model partitioning points through a deep reinforcement learning-based policy iteration algorithm, achieving multi-objective optimization of latency and energy consumption in complex environments. To validate the effectiveness of RLAMP, we conducted experiments using real-world datasets and compared its performance with existing model partitioning algorithms. The results demonstrate that RLAMP not only significantly reduces the model’s inference latency compared to other baseline methods but also lowers the device’s energy consumption.</p>