Small Language Models for Privacy-Aware On- Device IoT Intelligence: A Simulated Latency, Memory, and Accuracy Evaluation

This article presents a clearly labeled simulated empirical study in Computer Science and Engineering. A simulated IoT assistant workload compared cloud LLM inference with 3B-, 1.5B-, and 0.8B-parameter small language models running on edge-class hardware. Quantization, RAM use, time-to-first-token, tokens per second, energy per query, task accuracy, and offline availability were modeled across home automation, maintenance-query, and local- document tasks. The 1.5B 4-bit model produced the strongest simulated balance between quality and deployability, using 64% less memory than the 3B 8-bit model while retaining 93% of its task score. Fully local inference eliminated network dependence and reduced median response latency for short prompts. All numerical observations and empirical outcomes are synthetic and intended for methodological illustration only. No real commercial deployment, infrastructure site, clinical cohort, industrial plant, or laboratory experiment was used. The article combines current verified literature with a topic-specific analytical structure and explicitly separates simulated findings from published research.