The detection of hate speech in Arabic dialects presents unique challenges due to linguistic diversity and the scarcity of fine-tuned resources. This study provides a comparative analysis of several state-of-the-art Arabic language models, evaluating their performance in detecting hate speech across four datasets. These include two general dialectal datasets and two Maghrebi-specific datasets, focusing on Algerian and Tunisian dialects. By fine-tuning the models on standardized datasets, we assess their adaptability to the diverse and nuanced expressions commonly found in online hate speech. Our findings reveal significant performance disparities among models, with Maghrebi-specific models excelling in regional tasks but underperforming in multi-dialectal settings. These results emphasize the importance of diverse and context-rich pre-training data for effective hate speech detection in Arabic. This work provides valuable insights into the strengths and limitations of existing models and highlights pathways for improving Arabic NLP in low-resource and linguistically diverse contexts.