Accurate traffic flow forecasting is pivotal for transportation systems. However, existing deep learning methods, including Graph Neural Networks (GNNs) and Transformers, often struggle with several challenges: 1) Efficiently modeling long-range spatio-temporal dependencies, hindered by inherent GNN limitations or the quadratic computational complexity (O(N 2)) of Transformers; 2) Adaptively capturing dynamic spatio-temporal heterogeneity inherent in real-world traffic; 3) Effectively disentangling the complex superposition of mixed traffic patterns and multi-source influences. To address these limitations, we propose the Spatio-Temporal Selective State Expert Network (S3E-Net). This novel framework uniquely integrates a Mamba-inspired backbone, enabling efficient processing of long sequences with linear complexity, and a dynamic multi-expert mechanism, termed Dynamic Spatio-Temporal Expert Routing (DynSTER). DynSTER employs specialized experts to adaptively model diverse traffic phenomena. Extensive experiments on benchmark datasets demonstrate that S3E-Net achieves state-of-the-art performance, significantly outperforming existing methods in prediction accuracy and robustness across various horizons.