Financial sentiment analysis involves interpreting information from financial articles, news and social media to understand market trends and guide investment decisions. Numerous studies have used AI techniques, particularly large language models, to improve sentiment classification in the financial domain. Although these approaches have shown promising results, existing LLMs face two critical challenges: their large size, which results in high inference times and significant resource consumption, and their inability to fully account for the unique characteristics of financial sentiment classification. Unlike general sentiment analysis, financial sentiment requires robust reasoning and domain-specific expertise, which current models often do not adequately incorporate. To address these issues, this study proposes a lightweight chain-of-thought-based model for financial sentiment classification. We designed zero-shot chain-of-thought prompts tailored specifically for financial sentiment, integrating contextual reasoning into the classification process. This approach effectively improves accuracy by 2%–5% on financial sentiment classification tasks. In addition, we applied an activation-aware weight (AWQ) quantization algorithm to selectively reduce the model parameters. Experimental results demonstrate that our model significantly reduces memory usage by 90.3% and inference time by 87.9% compared to the baseline model, with only minimal impact on accuracy.