The dominance of English and Latin-based LLMs has created a gap in native Arabic models. These models are often hindered by the inclusion of English tokens, hampering effectiveness due to Arabic’s complex morphology and syntax. To address this deficiency, we introduce ArabianGPT, a series of transformer-based models specifically designed for Arabic. This series includes three variants of increasing size: ArabianGPT-0.1B, ArabianGPT-0.3B, and ArabianGPT-0.8B. A key contribution of these models is the development and integration of a purely Arabic tokenizer (AraNizer), which ensures precise and dedicated processing of Arabic text by excluding all non-Arabic tokens. Comparative analysis shows that fine-tuning greatly enhances performance in tasks such as question answering, and summarization. This highlights the efficacy of tailored transformer architectures in advancing Arabic NLP. Evaluating ArabianGPT against AraGPT2, GPT2, Bloom-7B1, and Llama-7B on benchmarks for scientific reasoning, common-sense understanding, cross-domain knowledge, and truthfulness, revealed promising results. ArabianGPT-0.8B achieved the highest score in HellaSwag (43.3) and the highest overall score (36.2), while ArabianGPT-0.3B excelled in TruthfulQA with a score of 52.5. Furthermore, fine-tuning these models on tasks like sentiment analysis and summarization led to significant improvements, with ArabianGPT-0.1B achieving 95% accuracy in sentiment analysis, despite its tiny size, and ArabianGPT-0.3B achieving 100%.