Micro-expression spotting in long videos aims to accurately detect subtle facial muscle movements of extremely short duration. To address the limitations of existing methods in coping with background noise, modeling long-range temporal dependencies, and handling the scale diversity of micro-expressions, a micro-expression spotting framework based on multi-scale Mamba is proposed. First, an improved feature pyramid network is introduced. It uses bilinear interpolation to optimize the fusion and alignment of multi-scale features and incorporates a convolutional block attention module (CBAM) to refine lateral connections and extract more discriminative temporal features. Second, a multi-scale Mamba adapter is designed to capture long-range temporal information through its selective state-space mechanism while maintaining low computational complexity. It is further combined with positional encoding to enhance the model’s ability to detect onset and offset boundaries. Finally, a decoupled detection head is employed for classification and regression, and varifocal loss and distribution focal loss are incorporated to optimize the two tasks jointly. Experimental results demonstrate that the proposed method achieves an F1-score of 0.3025 on SAMM-LV dataset, representing a 102.3% improvement over the baseline model. On the smaller CAS(ME)² dataset, the model achieves a competitive F1-score of 0.2188 and outperforms most mainstream deep learning approaches. These results validate the effectiveness and robustness of the proposed model in spotting micro-expressions within complex long video scenarios.