Abstract

Robust sentiment analysis in Portuguese is central to applications across Lusophone contexts, yet systematic evaluations still focus predominantly on English and proprietary systems. This paper presents a comparative study of 29 open-source Large Language Models (LLMs) and two proprietary models on Portuguese sentiment classification under four prompting strategies: Zero-Shot, Few-Shot, Chain-of-Thought (CoT), and CoT with Few-Shot (CoT+FS). Experiments on a unified three-class benchmark built from three public review corpora (about 3,000 instances) comprise roughly 372,000 inferences, totaling approximately 150M input tokens and 65M output tokens. Results show that CoT+FS generally yields the best performance for larger models, while several compact open-source models obtain competitive F1-scores with substantially lower computational cost, making them suitable for real-world deployments. We identify concrete teacher–student configurations tailored for knowledge distillation in Portuguese sentiment analysis.