Small Language Models for Edge and Mobile Computing: A Comparative Study of Accuracy, Latency, Energy Efficiency, and Privacy
IOSR Journal of Mobile Computing & Application · 2026-05 · 13 (5) · 23-27
DOI: 10.9790/0050-13052327 ↗Kurz gefasst
Fachartikel in der Zeitschrift IOSR Journal of Mobile Computing & Application, 2026-05. Verfasst von Dr. Pankaj Kumar.
Abstract (Originalsprache)
The rapid growth of generative artificial intelligence has traditionally been associated with large language models (LLMs) that require substantial computational resources, memory, and cloud infrastructure. Although these models provide strong language understanding and generation capabilities, their dependence on remote servers creates challenges related to latency, network connectivity, operational cost, energy consumption, and data privacy. Small Language Models (SLMs) provide an alternative approach by reducing model size and computational requirements while retaining useful language capabilities. This makes them particularly suitable for smartphones, tablets, Internet of Things (IoT) devices, embedded systems, and edge-computing platforms. Recent research has demonstrated that relatively compact models can provide competitive performance while being capable of local deployment. For example, Microsoft's Phi-3 Mini contains approximately 3.8 billion parameters and was specifically developed with local deployment, including smartphone execution, in mind [1]. Recent surveys also identify quantization, pruning, knowledge distillation, and architecture optimization as important techniques for enabling language models on resource-constrained devices [2,3]. This paper presents a comparative study of SLM deployment from four major perspectives: accuracy, latency, energy efficiency, and privacy. It reviews the technological foundations of edge-based language intelligence, examines modelcompression techniques, and proposes an evaluation framework for comparing SLMs with larger cloud-based models. The analysis indicates that the most appropriate model for edge computing should not be selected on accuracy alone; instead, a balanced evaluation involving response quality, inference time, energy consumption, memory requirements, and privacy should be used. The paper concludes with research challenges and future opportunities for efficient, secure, and sustainable on-device intelligence.
Zitieren
APA
Kumar, D. P. (2026). Small Language Models for Edge and Mobile Computing: A Comparative Study of Accuracy, Latency, Energy Efficiency, and Privacy. IOSR Journal of Mobile Computing & Application, 13(5), 23-27. https://doi.org/10.9790/0050-13052327
Vancouver
Kumar DP. Small Language Models for Edge and Mobile Computing: A Comparative Study of Accuracy, Latency, Energy Efficiency, and Privacy. IOSR Journal of Mobile Computing & Application. 2026;13(5):23-27. doi:10.9790/0050-13052327
BibTeX
@article{Kumar2026,
author = {Kumar, Dr. Pankaj},
title = {Small Language Models for Edge and Mobile Computing: A Comparative Study of Accuracy, Latency, Energy Efficiency, and Privacy},
journal = {IOSR Journal of Mobile Computing & Application},
year = {2026},
volume = {13},
number = {5},
pages = {23--27},
doi = {10.9790/0050-13052327}
}
EUVIMED