TII’s Latest Falcon-H1 Tops the Open Arabic LLM Leaderboard
The Falcon-H1 Arabic family is available in 3B, 7B, and 34B parameter sizes, a spread designed to balance performance with deployability.
News
- Non-Oil Growth Powers UAE’s Stronger-Than-Expected 2026 Outlook
- Musk’s xAI Raises $20 Billion in Upsized Funding Round
- CES 2026 Day 2 Signals AI’s Shift From Platforms to Physical Systems
- Saudi Arabia Nears Universal Internet Adoption, GASTAT Survey Shows
- Saudi Arabia to Allow FDI From Feb 1; Already Home to 20 of the World’s Top 30 Banks
- Twitter and Pinterest Founders Unveil a New App to Counter Social Media’s Harms
The Technology Innovation Institute (TII), the applied research arm of Abu Dhabi’s Advanced Technology Research Council, has launched Falcon-H1 Arabic, a new large language model that marks an architectural and performance leap in Arabic-language AI.
Unlike earlier iterations in the Falcon family, Falcon-H1 Arabic is built on a hybrid Mamba–Transformer architecture, moving away from a purely Transformer-based design. Essentially, Transformers comprehend language by re-reading everything simultaneously, whereas Mamba processes text sequentially, utilizing memory to make long documents faster and more cost-effective to handle.
The shift appears to be paying off. According to TII’s benchmark results, the model now tops the Open Arabic LLM Leaderboard (OALL), outperforming larger systems across a broad range of Arabic understanding and reasoning tasks.
The Falcon-H1 Arabic family is available in 3B, 7B, and 34B parameter sizes, a spread designed to balance performance with deployability across different infrastructure constraints. On OALL, the 3B variant achieved an average score of 61.87%, placing it roughly ten points ahead of leading models in the 4B class. The 7B model achieved 71.47%, surpassing competing systems in the 10B range, while the 34B version scored 75.36%, outperforming models with more than double its parameter count, including several systems with over 70 B parameters.
Beyond leaderboard rankings, TII observes strong performance on more specialised benchmarks such as 3LM for STEM reasoning, ArabCulture for cultural and contextual understanding, and AraDice for dialect comprehension. These results highlight the fluency in Modern Standard Arabic as well as broader coverage across dialects, domains, and real-world use cases.
A key technical feature of Falcon-H1 Arabic is its extended context window, which can span up to 256,000 tokens. In practical terms, this allows the model to process and reason over extensive documents within a single interaction, without the fragmentation that typically accompanies shorter context limits.
The launch builds on earlier Falcon-Arabic releases, which highlighted persistent gaps in quality Arabic LLMs despite rapid global progress in generative AI. TII is positioning Falcon-H1 Arabic as infrastructure-grade AI for multiple sectors, including education, healthcare, governance, and enterprise.
Falcon models have ranked at or near the top of multiple regional and global benchmarks since 2023. The models are publicly accessible through TII’s Falcon playground.
