Lightweight image super-resolution (SR) is critical for deploying high-quality reconstruction models on resource-constrained devices, such as mobile phones, embedded systems, or IoT devices. While existing methods often rely on stacking small convolutions or complex attention mechanisms, they struggle to balance efficiency and performance. In this work, a lightweight SR network called RepDynaMixer is proposed. It integrates dynamic kernel fusion and structural reparameterization for mobile-friendly efficiency without compromising quality. Two key innovations are introduced in this approach: (1) Dynamic Kernel Fusion, where multi-branch depth-wise convolutions (3 \(\times \) 3, 5 \(\times \) 5, 7 \(\times \) 7) adaptively fuse features through learnable gates, enabling content-aware feature extraction; and (2) Structural Reparameterization, which merges these multi-branch operations into a single 7 \(\times \) 7 depth-wise convolution during deployment, significantly reducing computational overhead. Performance is further enhanced by an edge enhancement branch and an adaptive channel split-shuffle mechanism that improves feature diversity. Experiments on benchmark datasets demonstrate that RepDynaMixer achieves competitive PSNR (26.37 dB) and SSIM (0.7915) with only 0.264M parameters and 14.9G FLOPs for \(4\times \) SR. It outperforms existing lightweight models such as CARN-M and IMDN, while requiring only 40% of their computational cost. This efficiency makes the model particularly suitable for mobile devices and real-time applications, effectively bridging the gap between laboratory research and practical deployment.