<p>With the proliferation of video data in edge computing environments, efficient video recognition techniques for resource-constrained edge devices have become increasingly important. However, existing methods struggle with limited computational resources on such devices. To address this challenge, we propose CTAMNet (CNN–Transformer attention mobile network), an efficient and accurate video recognition model designed specifically for edge deployment. CTAMNet enhances CNN-based feature extraction by introducing the temporal shift module (TSM) and dynamic tanh (DyT) activation, enabling effective temporal modeling without additional parameters. We further incorporate an Agent Attention mechanism and enhancement–compression (EC) modules to improve classification performance. The temporal–spatial feature fusion (TSFF) module effectively integrates spatial features from CNN and temporal features from attention mechanisms, capturing both visual appearance and temporal dynamics. Additionally, we introduce an adaptive focus loss function to enhance model training. Extensive experiments on UCF-101, Something–Something V1, and Something–Something V2 datasets demonstrate that CTAMNet achieves state-of-the-art accuracy while maintaining low computational overhead. The lightweight design of CTAMNet makes it readily deployable on resource-constrained edge devices, providing an effective solution for real-time video recognition tasks.</p>