Small Language Models For Edge Ai: A Review Of Efficient Architectures, Compression Techniques And On-Device Deployment

Small Language Models (SLMs) bring language understanding and generation directly to resource-constrained devices such as smartphones, wearables, IoT nodes and embedded systems. Large language models demand substantial memory, energy and connectivity, which limits their use in latency-sensitive, privacy-critical and offline settings. This paper reviews SLMs for Edge AI. It examines how compute-optimal scaling and high-quality training data enable compact models to perform well, and surveys contemporary open SLM families. Compression techniques are analysed in depth, including knowledge distillation, post-training quantization to four and eight bits, activation-aware and outlier-smoothing quantization, one-shot pruning, and parameter-efficient fine-tuning with low-rank adapters. Inference optimizations such as memory-efficient attention and speculative decoding are discussed. Trade-offs among accuracy, model size, inference latency and power consumption are compared. A lightweight deployment pipeline is proposed that selects a base model, adapts it to the task, compresses it and benchmarks it on target hardware before deployment. Research gaps, objectives and challenges including reasoning limits, hallucination, evaluation and secure updates are discussed. The review concludes that SLMs make private, low-latency, offline-capable language intelligence practical at the edge.