A comparative study of classical machine learning methods and transformer-based models for fake news detection on mobile devices is presented. Experiments conducted on the WELFake dataset indicate that transformer-based models achieve state-of-the-art classification performance, significantly outperforming classical approaches in terms of accuracy. However, this performance improvement is accompanied by higher computational and storage requirements. In contrast, classical models, such as SVM, deliver competitive performance while achieving substantially lower inference times and a reduced model footprint. Notably, lightweight transformers, such as MobileBERT, provide a practical balance by offering robust detection capabilities while maintaining feasibility for on-device deployment. The findings underscore the inherent trade-offs between predictive performance and operational efficiency, thereby emphasizing the need for tailored solutions in resource-constrained environments.