<p>Deep neural networks have revolutionized various domains with their exceptional performance; however, their large model sizes and high computational demands pose significant challenges for deployment in real-time and low-power applications. Neural network compression seeks to address these issues by reducing model complexity while maintaining accuracy. Optimizing compression strategies remains difficult due to the vast search space and the varying sensitivity of network layers to different compression techniques. In this paper, we propose a novel hybrid compression framework that combines structured compression via Singular Value Decomposition (SVD) with unstructured pruning, optimized through metaheuristic algorithms. By leveraging metaheuristics, our approach efficiently navigates the large, non-convex search space to automatically determine optimal, layer-specific compression parameters. This adaptive method balances the trade-off between compression ratio and model performance, tailoring the compression strategy to the unique characteristics of each layer. Extensive experiments across various neural network architectures and datasets demonstrate that our framework achieves significant model compression with minimal loss in accuracy. The proposed method offers a flexible and effective solution for deploying deep learning models in resource-constrained environments, facilitating practical applications in real-time and low-power settings.</p>