Model Compression for Sustainable AI in xG Wireless Networks: Recent Advances, Challenges, and Future Directions

Next-generation (xG) wireless systems, including sixth-generation (6G) networks and beyond, are expected to deliver data rates on the order of terabits per second and sub-millisecond latency. Meeting these requirements increasingly relies on artificial intelligence (AI)-enabled radio access and physical-layer (PHY) processing. However, realizing such AI-driven PHY functionality with deep learning (DL) is challenging because deep neural networks (DNNs) are computationally intensive and memory hungry, often exceeding the capabilities of resource-constrained user equipment (UE) and edge hardware. This paper surveys model-compression techniques for efficient wireless intelligence, focusing on pruning, quantization, and knowledge distillation (KD), together with architectural and algorithmic optimizations. For each technique, we summarize theoretical foundations, practical implementation strategies, and wireless-specific considerations, and discuss how design choices translate into latency, energy, and memory outcomes on deployment hardware. We review applications across core PHY wireless tasks, including automatic modulation classification (AMC), channel state information (CSI) processing and feedback, beamforming (BF), recognition and identification, channel estimation and detection, and localization. Drawing on comparative analysis of more than 50 studies, we highlight trade-offs among model size, computational complexity, energy consumption, and task-level performance under wireless evaluation protocols. We further discuss hardware–software compatibility considerations for compressed model deployment and outline open challenges and future research directions for compression-aware deployment in xG wireless systems.