Sleep spindles are critical biomarkers for sleep scoring and play a significant role in assessing cognitive abilities as well as related disorders. However, traditional methods relying on manual annotation by sleep experts are time-consuming and labor-intensive. Therefore, the development of automated sleep spindle detection algorithms has important clinical significance. Nevertheless, this task is still challenging due to the limited datasets and the variability of inter-expert annotation. In this study, we propose a Mamba-based attention-Unet framework, referred to as Spindle-UMamba, to detect sleep spindle effectively. The U-Net architecture facilitates the efficient utilization of limited EEG samples by the model, while Mamba enhances the model’s capacity to capture fine-grained details and contextual information embedded within EEG signals. Additionally, an attention mechanism is also integrated into the model to rescale the features selectively, thereby enhancing the performance of model. Finally, the proposed method was evaluated on the widely used benchmark dataset of MASS. Spindle-UMamba achieves state-of-the-art in the MASS dataset with average F1 scores of 80.0% and 79.9% when using annotations from two independent experts as ground truth. Compared to previously proposed advanced methods, our model achieved higher and more balanced performance, demonstrating its robustness and effectiveness in sleep spindle detection.