Abu Dhabi’s Technology Innovation Institute (TII) has unveiled Falcon-H1 Arabic, a new large language model designed specifically for Arabic. The model’s largest 34-billion-parameter version has secured the top spot on the Open Arabic LLM Leaderboard (OALL), outperforming bigger international models such as Meta’s Llama-3.3 (70B) and Alibaba’s Qwen2.5 (72B) in Arabic comprehension, reasoning, and dialect recognition. Falcon-H1 Arabic is available in three sizes: 3B, 7B, and 34B parameters. The 3B model outperforms Microsoft’s Phi-4 Mini on Arabic tasks, while the 7B ranks among the strongest mid-size Arabic models. The flagship 34B model shows higher accuracy across complex language tasks, including handling long texts and cultural nuances. Unlike other AI systems adapted from English, Falcon-H1 was built with an Arabic-first approach using a special Mamba-Transformer architecture. This helps the model handle Arabic’s challenging grammar, sentence structures, and regional dialects consistently. The new model can also perform well in STEM-related reasoning, suggesting it can do more than just language tasks. Its relatively smaller size means it needs less computing power, making it cheaper and faster for local businesses and governments to use. Practical uses include analyzing long legal and medical documents, improving Arabic-language education tools, and powering government customer service bots that understand various Arabic dialects. The launch fits into the UAE’s wider plan to develop its own AI that truly understands regional language and culture. H.E. Faisal Al Bannai, Adviser to the UAE President, said Falcon-H1 Arabic shows the UAE’s commitment to responsible AI development. TII CEO Dr. Najwa Aaraj said the model fills gaps where existing AI tools still struggle with Arabic. Officials say Falcon-H1 is not meant to replace global AI systems but to work alongside them, offering better Arabic-language support for users in the region.