Large Language Models (LLMs) have demonstrated impressive capabilities in various natural language processing tasks across different languages. However, their performance in low-resource languages and dialects, such as Moroccan Arabic (MA), requires further investigation. This study evaluates the performance of ChatGPT-4, different fine-tuned BERT models, FastText as text representation, and traditional machine learning models on MA sentiment analysis. Experiments were done on two open source MA datasets: an X(Twitter) Moroccan Arabic corpus (MAC) and a Moroccan Arabic YouTube corpus (MYC) datasets to assess their capabilities on sentiment text classification. We compare the performance of fully fine-tuned and pre-trained Arabic BERT-based models with ChatGPT-4 in zero-shot settings.