Time series forecasting (TSF) is fundamental to data-driven decision systems. Most existing methods treat historical observations as equally important, overlooking the fact that temporal dynamics become more complex near the prediction horizon. This uniform assumption may lead to redundant modeling and insufficient focus on critical periods. To address this, we propose NuTS, a framework for Non-uniform Sampling with Distance-Dependent Adaptive Fusion. NuTS introduces Distance-Stratified Temporal Downsampling (DSTD) to perform non-uniform sampling based on temporal distance, enabling near-dense and far-sparse modeling of different temporal patterns. In addition, a Dual-Path Mixer (DPM) jointly captures temporal dynamics and inter-channel dependencies, where channel dependencies serve as auxiliary gating to suppress redundant noise. Finally, we introduce Distance-Dependent Adaptive Fusion (DDAF), which employs a gated Mixture-of-Experts mechanism to adaptively fuse representations across temporal distances. Extensive experiments on benchmark datasets demonstrate that NuTS consistently outperforms state-of-the-art methods.