Long-term time series forecasting plays a critical role in numerous real-world applications. Although recent deep learning models have shown strong performance in either the time domain or the frequency domain, they still encounter two major challenges: First, single-domain modeling fails to capture complementary signals, with time-domain models obscuring the long-term cycles and frequency-domain approaches smoothing out the localized variations. Second, existing time-frequency hybrid methods are overly simple and forceful, leading to fused representations that contain redundancy and conflicts. To address these challenges, we propose TFMoE, a Time-Frequency Mixture-of-Experts framework inspired by the flexibility of Mixture-of-Experts. Concretely, we design two complementary expert types: the Time-Domain Expert (TDE), which models temporal dependencies and trends, and the Frequency-Domain Expert (FDE), which captures cyclicity and spectral correlations through complex-valued mixing. Crucially, TFMoE dynamically activates multi-scale and domain-specific experts via adaptive pathways, enabling specialized modeling for both temporal and spectral patterns. Through hierarchical cross-domain interaction across layers, TFMoE progressively refines temporal and spectral representations, achieving balanced integration. Extensive experiments on seven real-world datasets show that TFMoE consistently achieves superior forecasting performance and robustness compared to state-of-the-art models.