Why teaching Arabic to AI is hard and how UAE researchers are solving it

Building artificial intelligence that understands Arabic dialects requires solving a problem most global tech companies haven’t attempted: training a single system to handle Egyptian, Levantine, Gulf, and Maghrebi Arabic alongside modern standard Arabic — without losing quality.

Abu Dhabi’s Technology Innovation Institute (TII) achieved this with Falcon-H1 Arabic by expanding training data beyond formal written Arabic to include dialectal sources, then filtering to ensure linguistic diversity across regions.

“Falcon-H1 Arabic was trained to handle a broad range of widely used dialects,” Hakim Hacid, chief researcher of TII’s Artificial Intelligence and Digital Science Research Center, told Khaleej Times, noting: “The training data was intentionally expanded beyond formal written Arabic to include dialectal sources, and carefully filtered to ensure linguistic diversity.”

The technical challenge stems from Arabic’s structure. “It combines rich morphology, flexible sentence structure, and high variation between Modern Standard Arabic and regional dialects,” Hacid added.

Most global AI systems treat dialects as variations of a single language, applying the same processing approach used for English. This fails because dialectal Arabic involves different vocabulary, grammatical structures, and pronunciation patterns that fundamentally change meaning.

The solution required architectural innovation. Falcon-H1 Arabic uses a hybrid system combining transformer attention with state space models called Mamba. “This allows the model to process information more efficiently, particularly over long sequences, while maintaining strong reasoning capabilities,” Hacid said.

This efficiency explains why a 34-billion-parameter model outperforms systems with 70+ billion parameters. “Performance is not just about scale,” Hacid noted. “Combined with improvements in data quality, dialect coverage, and optimisation, this enables a smaller model to outperform larger models on Arabic benchmarks.”

The model’s 256,000-token context window enables analyzing entire documents while maintaining coherence. “Users can analyse entire legal cases, medical histories, or research papers at once,” Hacid explained. “This was previously impractical for Arabic AI systems.”

Applications include analysing legal documents without translation, summarising medical records mixing formal and dialectal language, and powering enterprise systems operating natively in Arabic.

Arabic preservation

TII’s results tell us that advanced AI research extends beyond a few countries. “Falcon’s performance shows that teams based in the UAE are contributing meaningful architectural innovation and building models that compete at the highest global level,” Hacid said.

The development aligns with goals around Arabic preservation in technology. “By prioritising Arabic language support, including dialects, the work aligns technology development with cultural and linguistic realities,” Hacid said, giving users “the ability to educate, work, and enjoy the cyber world in their native language.”

However, significant gaps remain. Hacid outlined three priority areas for future development: integrating more dialects, particularly those with limited digital resources; achieving full parity with English capabilities, including advanced reasoning tasks; and entering multimodal AI that combines text, images, and speech—all natively in Arabic rather than through translation layers.

“Enabling the integration of more dialects is important to continue the effort of preserving low resources dialects,” Hacid said. “Beyond generation, Arabic must also enter the AI space fully as a first class citizen. This means that all capabilities offered by, for example, English, should be also provided in Arabic, in a native way.”

The model’s expandability matters for this roadmap. “The model is expandable to more languages without losing the existing quality in the current dialects, following the right finetuning recipe,” Hacid noted, suggesting future versions could incorporate additional regional variations without degrading performance on existing ones.

Falcon-H1 Arabic’s release as an open-source model accelerates this development by allowing researchers, developers, and institutions across Arabic-speaking regions to adapt and extend the technology for specific dialects, industries, or applications — building toward the goal of Arabic as a “first class citizen” in AI rather than a translated afterthought.