Single-cell multi-omics clustering can reveal cell heterogeneity, providing critical support for precision medicine. Existing methods utilize autoencoders (AE) to reduce dimensionality and reconstruct multi-omics data, combining graph autoencoders (GAE) to capture topological relationships between cells, extracting feature representations for single-cell multi-omics and effectively improving clustering performance. However, we found that the introduction of AE may lead to feature redundancy, potentially impairing the topological modeling capability of GAE. Thus, we propose the single-cell multi-order topological cross-omics embedding fusion method (SMTFusion). Specifically, SMTFusion constructs cell graphs with multi-order topological structures, which are based on multi-omics data, to enhance information propagation among multi-hop neighboring cells. It employs self-attention and cross-omics attention mechanisms to enhance cell feature representations across different omics, while adaptively fusing multi-order topological features with learnable weight coefficients. Finally, SMTFusion optimizes the feature representation of single-cell multi-omics by jointly leveraging reconstruction loss, alignment loss, and dual supervised clustering loss. We validated SMTFusion on four datasets, and experimental results demonstrate that the proposed method exhibits superior performance in clustering accuracy and robustness.