TII announces new Arabic AI model family

Falcon-H1 Arabic tops Open Arabic LLM Leaderboard

#UAE #LLMs - Technology Innovation Institute (TII), the applied research arm of Abu Dhabi's Advanced Technology Research Council (ATRC), has announced Falcon-H1 Arabic, a high-performing Arabic large language model family. Built on a hybrid Mamba-Transformer architecture, the models are available in 3 billion, 7 billion and 34 billion parameter sizes. According to TII’s evaluation using the Open Arabic LLM Leaderboard (OALL), the new models outperform systems with over 70 billion parameters including Meta’s Llama-3.3 70B and Alibaba Cloud’s Qwen2.5 72B. The new models aim to provide Arabic language AI capabilities with improved data quality, dialect coverage and long-context processing up to 256,000 tokens.

SO WHAT? - The new Falcon-H1 Arabic model family addresses the scarcity of high-performance Arabic language AI systems, providing tools optimised for the linguistic and cultural requirements of Arabic-speaking regions. By delivering competitive performance at smaller parameter sizes, the models enable organisations with limited computational resources to deploy capable Arabic AI applications across education, healthcare, governance and enterprise sectors.

Here are some key facts about the new Falcon H1 Arabic model family:

• Abu Dhabi-based Technology Innovation Institute (TII) has announced Falcon-H1 Arabic, a high-performing Arabic large language model family. Representing a complete departure from previous transformer-based Falcon 3 versions, the new LLMs are built on a hybrid Mamba-Transformer architecture.

• The Falcon-H1 Arabic family includes three model sizes: Falcon-H1 Arabic 3B; Falcon-H1 Arabic 10B; and Falcon-H1 Arabic 34B. Licencing for the new models has not yet been announced by TII.

• According TII the new Arabic models perform as follows using the Open Arabic LLM Leaderboard (OALL) to evaluate them:

  • Falcon-H1 Arabic 3B scores an average of 61.87 percent, 10 points ahead of leading 4B competitors, such as Microsoft’s Phi-4 Mini.

  • Falcon-H1 Arabic 10B scores an average of 71.47 percent, surpassing all ~10B models, including Qatar’s Fanar-1-9B and Saudi Arabia’s HUMAIN ALLaM 7B model.

  • Falcon-H1 Arabic 34B scores 75.36 percent, outperforming even 70B+ parameter systems, including Alibaba Cloud’s Qwen2.5 72B and META’s Llama-3.3 70B, demonstrating efficiency gains from the hybrid Mamba-Transformer architecture.

• The models extend context length capabilities to 256,000 tokens, enabling analysis of lengthy documents including legal contracts, medical records and academic papers whilst maintaining contextual continuity across the entire interaction.

• The new Arabic model family builds on years of foundational Arabic AI work, advancing architecture, data quality and long-context reasoning to create capabilities for education, healthcare, governance and enterprise applications.

• According to TII, the models achieve strong results on specialised benchmarks including 3LM for STEM reasoning, ArabCulture for cultural understanding and AraDice for dialect comprehension, demonstrating linguistic depth beyond general language tasks.

ZOOM OUT - The Falcon-H1 Arabic release follows last May’s launch of Falcon 3 Arabic, the first Arabic language model in the Falcon family, but is built on a hybrid Mamba-Transformer architecture. The Falcon 3 Arabic 7B was trained on 600 giga tokens of Arabic, multilingual and technical data spanning Modern Standard Arabic and regional dialects, outperforming regionally available Arabic models and matching systems up to 10 times its size on Open Arabic LLM Leaderboard benchmarks. The Falcon model series, first released in 2023, has been downloaded over 55 million times globally, establishing TII's position in open-source AI development beyond traditional North American and European centres.

[Written and edited with the assistance of AI]

Note: Falcon-H1 Arabic models are not visible on the live OALL

LINKS

• Falcon chat playground (TII)

• Falcon H1 Family Technical Report (arXiv)

• Falcon H1 Family code (GitHub)

Read more about Falcon large language models:

• TII releases compact reasoning model Falcon-H1R (Middle East AI News)

• Falcon 3 LLM series gets first Arabic model (Middle East AI News)