Introducing Falcon-H1-Arabic, a groundbreaking advancement in Arabic language AI, marking a major leap in both architectural innovation and real-world performance. This new model family, developed by the Technology Innovation Institute (TII), represents the culmination of extensive research, community-driven feedback, and technical refinement, delivering three powerful variants—3B, 7B, and 34B—that set new benchmarks for Arabic natural language processing. Building on the foundation of Falcon-Arabic, which was launched earlier and widely adopted across the Arab world, the team identified key challenges including long-context understanding, dialectal variation, mathematical reasoning, and domain-specific knowledge. Rather than making incremental upgrades, they reimagined the architecture entirely, introducing the Falcon-H1 hybrid design—a first for Arabic language modeling. At its core, Falcon-H1-Arabic integrates State Space Models (SSMs), specifically Mamba, with Transformer attention mechanisms within each model block. Both components operate in parallel and their outputs are fused before the block’s final projection. This hybrid approach combines the linear-time scalability of Mamba—ideal for handling extremely long sequences—with the precise long-range modeling strengths of Transformers. This is particularly beneficial for Arabic, given its rich morphology and flexible syntax, enabling superior coherence and reasoning across lengthy texts. Context windows have been dramatically expanded: the 3B model supports up to 128K tokens, while the 7B and 34B versions handle 256K tokens—equivalent to hundreds of pages of text or entire novels. This enables transformative applications in legal analysis, medical documentation, academic research, and sustained conversational AI. The models are trained to effectively utilize their full context, overcoming the common "lost in the middle" issue seen in earlier systems. The pre-training data pipeline was rebuilt from the ground up using deep linguistic analysis tailored to Arabic orthography, diacritics, morphology, and syntax. This rigorous filtering process removed noise and ensured high-quality, stylistically consistent content. Dialect coverage was significantly enhanced to include Egyptian, Levantine, Gulf, and Maghrebi varieties, ensuring the models understand and generate authentic regional Arabic. Despite this focus on dialects, the models retain strong multilingual capabilities through balanced training on Arabic, English, and multilingual content totaling around 300 billion tokens—supporting robust performance in STEM, code, and cross-lingual tasks. Post-training follows a two-stage process: supervised fine-tuning (SFT) with high-quality Arabic instructions, long-context examples, and reasoning tasks, followed by direct preference optimization (DPO) to refine alignment, conversational quality, and consistency. This ensures the models not only process long inputs but also maintain coherence, avoid drift, and respond helpfully across multi-turn interactions. Benchmark results confirm Falcon-H1-Arabic’s leadership. On the Open Arabic LLM Leaderboard (OALL), all three models outperform state-of-the-art models of similar or larger sizes. The 3B model scores around 62% on OALL, surpassing models like Gemma-4B and Phi-4-mini by over ten points. The 7B model achieves 71.7% on OALL, outperforming all ~10B-class models. The 34B model reaches approximately 75% on OALL, exceeding even large models like Llama-3.3-70B and AceGPT2-32B—demonstrating the power of the hybrid architecture. On specialized benchmarks, Falcon-H1-Arabic excels: 92% on the native split of 3LM (Arabic STEM), 80% on ArabCulture, and over 50% on AraDice for dialect coverage. These results reflect real-world readiness across diverse tasks. Deployment scenarios are well-supported: the 3B model is ideal for edge devices, agentic systems, and high-throughput applications; the 7B model serves as a versatile production-grade solution for chatbots, summarization, and enterprise tools; the 34B model is designed for high-stakes domains like legal and medical analysis, where accuracy and long-context reasoning are critical. As with all AI models, Falcon-H1-Arabic may reflect training data biases and generate hallucinations. Outputs should not be used for critical decisions without expert validation. Performance may degrade at extreme context lengths, so task-specific evaluation is recommended before production use. The team acknowledges the contributions of the broader Arabic NLP community and TII colleagues for their vital support. Falcon-H1-Arabic is now available in three sizes on Hugging Face, representing a major milestone in making advanced Arabic AI accessible, capable, and responsible.