MOTIVATION: 5-Methylcytosine (5mC) plays an important role in gene regulation and development. Although nanopore sequencing has enabled direct detection of 5mC, existing methods still face several limitations, including poor generalization across species and sequence contexts (CpG/CHG/CHH), as well as suboptimal integration of sequence and current signals. RESULTS: Here, we present MethyNano, a deep learning framework incorporating a contrastive learning strategy to detect 5mC from nanopore reads. By encouraging more discriminative and stable representations, the contrastive objective improves the model's sensitivity to rare sequence contexts and reduces its prediction uncertainty in challenging regions. Across datasets from Arabidopsis thaliana, Oryza sativa, and Homo sapiens, our model achieves superior performance on key metrics compared with other existing methods. Extensive cross-species and cross-motif experiments demonstrate the robust generalization performance of MethyNano, while dimensionality-reduction visualizations of learned features provide an intuitive view of the model's efficient representation capability. Moreover, our ablation studies show that MethyNano's architecture enables more effective integration of critical features, leading to higher predictive accuracy. AVAILABILITY AND IMPLEMENTATION: The project code is available at https://github.com/baigeHUI/MethyNano and https://doi.org/10.5281/zenodo.19858400.