Sentiment analysis is a primary task in Natural Language Processing (NLP), where the goal is to classify emotions from text. This paper investigates and compares two prominent approaches for Arabic sentiment analysis: encoder-based models, which utilize Transformer Embeddings such as AraBERT and CAMeLBERT, and encoder-decoder-based models, which involve the Fine-Tuning of large language Models (LLMs) like ArabianGPT-0.1B and AraGPT2. We explore how each approach handles the complexities of Arabic morphology and semantics to capture nuanced sentiment in the text. Empirical evaluation on LABR dataset (a large-scale Arabic Book Reviews dataset) shows that the Fine-Tuning of LLMs consistently outperforms the Embedding-based models offering superior accuracy and generalizability. These findings underscore the potential of Fine-tuned LLMs for real-world applications, including social media monitoring, opinion mining, and customer feedback analysis in Arabic-speaking contexts.