TII launches an AI model built for Emirati Arabic

Abu Dhabi's Technology Innovation Institute (TII) launched its new Falcon-Emirati model this week, trained specifically on the Emirati dialect of Arabic. The 7-bn parameter model is part of a push to make Arabic-language models more useful by training them on idioms and slang rather than formal language.

Why it matters: Arabic AI models have historically suffered from limited data, so many models were built on English data translated into Arabic, Dr. Hakim Hacid (LinkedIn), chief researcher at TII’s AI and Digital Research Center, tells EnterpriseAM. That approach “resulted in losing all the specificities of the languages, of the culture,” he says. TII also released a multilingual speech-transcription model and an Arabic model for extracting text from images and documents.

How it scores: Falcon-Emirati scored 84.83% on an Emirati Arabic test, beating every open-source Arabic and multilingual model it was compared with, TII says. The institute designed the test itself, but also made its data public so outsiders can check the results. Closed models like ChatGPT and Claude weren’t part of the comparison.

Next: TII plans to build models for other Arabic dialects, but wants user feedback on the Emirati model first, Hacid says. He declined to give a timeline for when other models will arrive.

Other dialect models: Teams led by Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi’s AI research university, released Atlas-Chat, a model for Moroccan Arabic, in 2024, and Nile-Chat, for Egyptian Arabic, last year.