<p>Implicit hate speech, unlike explicit hate speech, is characterized by its subtle and nuanced nature, presenting significant challenges for automated detection systems. Previous studies have explored the complexities of identifying such content, emphasizing the role of cultural and linguistic context in decoding sarcasm, coded language, and implied meanings. Building on this foundation, this study focuses on the challenges and progress in detecting implicit hate speech in Arabic, an important yet overlooked area. It fills the gap in current detection systems by developing a framework designed specifically for implicit hate speech in Arabic. We present a novel Arabic implicit hate speech dataset of approximately 2500 records, collected from several hate speech datasets. Moreover, various innovative models were developed to identify implicit hate speech. Specifically, machine learning, deep learning, and transformer-based architectures and hybrid models were compiled. The results demonstrate that two hybrid models performed notably well: the first architecture, which combines three high-performing deep learning models (BiLSTM, CNN, and GRU), achieved an accuracy of 82%, while the second architecture, which combines the two pre-trained large language models MarBERT and Qarib, achieved an accuracy of 86%. Furthermore, GRU combined with FastText obtained an accuracy of 83%, and MarBERT alone reached 87%, indicating that transformer-based models tend to outperform traditional deep learning approaches in this task. Moreover, we examined misclassification patterns, emphasizing the importance of refining detection methods to account for all forms of implicit hate speech and leveraging attention mechanisms to better capture contextual subtleties.</p>