EVALUATING THE USABILITY AND EFFICIENCY OF COMPACT LARGE LANGUAGE MODELS FOR SENTIMENT ANALYSIS TASKS

The rapid proliferation of large language models (LLMs) has enabled substantial advances in natural language processing, but their resource requirements create barriers to accessibility, cost-efficiency, and sustainable deployment. This paper presents a systematic benchmark of compact LLMs—ranging from 135M to 8B parameters— on eight diverse sentiment analysis datasets, including binary, fine-grained, domain-specific, social media, and aspect-based tasks. Models were evaluated in a zero-shot setting using normalized accuracy metrics to ensure comparability across datasets of varying difficulty, alongside measurements of latency and memory usage.
Our findings reveal that mid-sized models in the 3–4B parameter range, particularly Gemma 3 4B and Microsoft Phi-4 Mini, consistently outperform or rival larger 7–8B models while offering significantly lower latency and memory footprints. Sub-1B models, while largely ineffective in zero-shot conditions, retain potential for high-throughput pipelines and fine-tuned deployments. Conversely, larger 7–8B models remain valuable for accuracy-critical tasks but incur diminishing returns relative to their computational cost. These results highlight the importance of balancing accuracy with efficiency and suggest that mid-sized models constitute the practical “sweet spot” for zero-shot sentiment analysis on commodity hardware.