The collaborative computing architecture of cloud-edge intelligent systems is subject to dynamic heterogeneity across devices, which poses significant challenges for efficient and adaptive model training and inference. Traditional static model splitting approaches typically lack the responsiveness to accommodate fluctuating resource availability, often resulting in suboptimal performance. To overcome these limitations, we formulate the model splitting as a constrained optimization problem that incorporates the heterogeneous resources of the edge devices. Then, we propose an adaptive model splitting framework, named AdaSplit, which dynamically adapts to heterogeneous edge nodes. The proposed framework comprises two key components: (1) an adaptive model splitting algorithm that determines partition ratios via local training time and then dynamically assigns layers to appropriate devices, wherein device heterogeneity is converted to training time; and (2) a hierarchical aggregation scheme that iteratively refines the global model, wherein the central server aggregates client updates in a layer-wise manner. Extensive discrete-event simulations, replicating practical cloud-edge intelligent environments, demonstrate the effectiveness of our approach. Experimental results show that AdaSplit consistently outperforms state-of-the-art baselines, reducing latency by at least 30% while maintaining accuracy comparable to full-model training under various heterogeneous conditions.