Deep neural networks, such as convolutional neural networks (CNNs) and vision transformers (ViTs), have significantly advanced unsupervised anomaly localization in industrial manufacturing. However, reconstruction-based methods still encounter two primary challenges: (1) overgeneralization due to insufficiently discriminative spatial information, and (2) the high computational complexity of ViT-based models. To this end, we propose the first spatial-frequency parallel mamba model (SFPM \(^{\textbf {2}}\) ), which jointly explores spatial and frequency correlations in industrial images. Specifically, SFPM \(^{\textbf {2}}\) incorporates a frequency-guided parallel adapter within the mamba structure to separately model long-range dependencies in both spatial and frequency domains. Leveraging the global receptive field and linear computational complexity of mamba, the model achieves strong adaptability to unseen domains. To effectively fuse the complementary features from both domains, an adaptive cross-attention fusion strategy is introduced. Furthermore, a fourier transform-based high-frequency enhancement block is designed to mitigate detail loss during decoding. Finally, a customized hybrid loss function guides model training and anomaly estimation. Experimental results on two large-scale public datasets demonstrate that SFPM \(^{\textbf {2}}\) achieves state-of-the-art performance in both quantitative and qualitative evaluations.