On-Device Large Language Models for Mobile Applications: A Systematic Survey of Optimization Strategies, Deployment Challenges, and Privacy Implications

The rapid advancement of large language models (LLMs) has accelerated a major transition from cloud-centric inference toward on-device deployment on smartphones and edge devices. Running LLMs locally offers important advantages, including lower latency, offline availability, reduced operational cost, and improved data privacy. However, practical deployment on mobile hardware remains challenging due to strict constraints involving memory bandwidth, thermal envelopes, battery capacity, heterogeneous accelerators, and runtime fragmentation. This survey presents a systematic and hardware-aware review of on-device LLM deployment for mobile systems, synthesizing recent research spanning model architecture, compression techniques, runtime ecosystems, benchmarking methodologies, and privacy considerations. We analyze the growing shift toward compact small language models (SLMs), activation-aware quantization methods such as AWQ, KV-cache optimization strategies, and heterogeneous scheduling approaches designed for modern mobile System-on-Chip (SoC) architectures. In addition, we compare widely used inference frameworks including llama.cpp, MLC LLM, ExecuTorch, ONNX Runtime Mobile, and TensorRT-LLM, highlighting their tradeoffs across portability, throughput, backend support, and deployment complexity. Beyond model efficiency, this survey emphasizes an important systems-level observation: sustained mobile inference is fundamentally constrained by memory bandwidth and thermal sustainability rather than peak computational throughput alone. To structure deployment tradeoffs, we introduce the ALPE framework, which evaluates mobile LLM deployment across four dimensions: Accuracy, Latency, Privacy, and Energy. Furthermore, we discuss unresolved challenges involving standardized benchmarking, vendor-specific NPU fragmentation, long-context inference, thermal-aware scheduling, and secure on-device agent execution. By integrating insights across hardware, runtime systems, and model optimization, this survey aims to provide a comprehensive foundation for future research in practical and sustainable on-device LLM deployment.