<p>Long-term time series forecasting (LTSF) encounters inherent challenges due to the intricate and dynamic characteristics of time series data, which frequently exhibit both short-term fluctuations and long-term dependencies. Fixed-resolution models struggle to capture multi-scale temporal patterns, and implicit multi-resolution models often smooth out critical high-frequency variations or obscure how different temporal scales contribute, limiting accuracy and interpretability. To address these limitations, this study proposes M-Mamba, a novel multi-resolution Mamba model designed specifically for LTSF, where high resolutions capture fine-grained local patterns, low resolutions capture broader trends, and adaptive fusion weights emphasize informative resolutions to improve forecasting accuracy and interpretability. An adaptive channel dropout strategy dynamically adjusts dropout rates for each resolution during independent processing, preventing over-reliance on specific channels while preserving model flexibility. Extensive experiments on eight real-world datasets show that M-Mamba achieves state-of-the-art accuracy with competitive efficiency, and a lightweight variant further improves the accuracy–efficiency trade-off. Given the computational cost of multi-resolution LTSF, training and inference are performed on GPUs, and the independent resolution branches are naturally amenable to parallel execution on multi-GPU systems, with potential to further increase throughput. These experimental results validate the effectiveness of the proposed multi-resolution strategy and adaptive channel dropout mechanism for long-term time series forecasting.</p>