Comparative Analysis of Mobile-Friendly Large Language Models: Performance, Efficiency, and Accuracy

This research presents a comparative analysis of various mobile-friendly LLMs, evaluating their performance in terms of inference speed (tokens per second) and response accuracy. The study benchmarks models such as Llama 3:2.1B, Gemma 2:2B, Deepseek R1:1.5B , and Qwen 2.5, testing them in an offline environment to determine their suitability for mobile deployment. The results highlight the trade-offs between model size, inference speed, accuracy, BERTScore, BLEU, and ROUGE-L, showing that smaller models achieve higher throughput but often compromise response quality. The findings provide valuable guidance for developers and enterprises seeking to implement LLM-based solutions on mobile devices while balancing computational efficiency and accuracy.