The rapid growth of China’s food delivery market has resulted in more than 14 million active riders; however, significant income disparities exist, with novice riders earning nearly half as much as experienced ones. Equipped with e-bikes, modern riders must optimize delivery routes under multiple objectives while monitoring battery levels and making timely swapping decisions to maintain efficiency. Current academic and industrial methods fail to address these coupled challenges. To bridge this gap, we propose ENDeliver, the first energy-aware framework for e-bike delivery route planning. ENDeliver consists of two core components: a large-scale energy-aware pretrained model with an EnergyAttention mechanism trained on millions of real-world delivery trajectories, which learns riders’ routing and battery-swapping experience to assist novice riders in improving performance; and an inference module that iteratively queries the pretrained model to generate next-step moving and swapping actions, thereby constructing efficient delivery routes under energy constraints. Experiments on multiple real-world delivery trajectory datasets show that ENDeliver produces shorter routes in 84% of cases, reduces average delivery distance by 42%, achieves 82.9% accuracy in emulating the timing and location of experienced riders’ battery-swapping decisions.