Abu Dhabi’s Technology Innovation Institute (TII), the applied research arm of the Advanced Technology Research Council (ATRC), on Tuesday released three artificial intelligence models built around how Arabic is spoken, heard and read in the UAE. The centrepiece, Falcon-Emirati, is a 7-billion-parameter language model trained to understand Emirati Arabic, including the dialect’s idioms, proverbs and poetry. It arrives with Falcon-ASR, a 1.6-billion-parameter speech recognition model, and Falcon-OCR-Arabic, a lightweight model that extracts Arabic text, tables and formulas from images and documents.
Dr. Najwa Aaraj, Chief Executive Officer of TII, tied the release to the institute’s sovereign AI agenda. “Sovereign AI capability must reflect the language used in daily life,” she said. “Emirati Arabic carries distinctive expressions, cultural references and ways of communicating. Falcon-Emirati and Falcon-ASR will help ensure that the next generation of AI understands not only Arabic but how Emirati communities actually speak it.”
Arabic AI has learnt the formal register far better than the spoken ones
Arabic is spoken by hundreds of millions of people across more than 20 countries, yet the Arabic most AI systems have absorbed is Modern Standard Arabic, the register of news bulletins, textbooks and official documents. Daily conversation in Dubai, Riyadh, Cairo or Casablanca happens in regional dialects that differ in vocabulary, grammar and rhythm, and those dialects appear far less often in the written material used to train large language models. TII describes the problem precisely: a model trained mainly on the formal register may understand every word of an Emirati sentence and still miss what the speaker meant, particularly when the meaning sits inside a proverb or a line of Nabati poetry.
That shortfall has practical consequences for the region’s governments and enterprises, which have spent recent years deploying chatbots, voice assistants and document automation in Arabic. A customer service agent that replies in textbook Arabic to a question asked in dialect feels stilted at best, and a transcription tool that mishears local speech produces records nobody can search.
Falcon-Emirati extends a two-year push to make Falcon fluent in Arabic
TII has been building the foundation for the new model in stages. The institute introduced Falcon Arabic, the first Arabic model in its Falcon series, in May 2025, and followed it in January 2026 with Falcon-H1-Arabic in 3-billion, 7-billion and 34-billion-parameter versions. Falcon-Emirati is built on Falcon-H1-Arabic. It was trained on native Emirati content, Modern Standard Arabic material about Emirati culture and heritage, and synthetic data generated with the help of Emirati linguistic resources.
TII reported that the model scored 84.83% on Alyah, a native Emirati Arabic benchmark covering everyday language, figurative expressions, heritage knowledge and poetry, and that it outperformed every Arabic and multilingual open-source model evaluated. The institute did not name the models it tested against, give their scores, or say who developed Alyah, so the figure cannot yet be checked independently. Falcon-Emirati will be offered through a Falcon Chat platform. TII has not said whether the model’s weights will be published for developers.
H.E. Faisal Al Bannai, Adviser to the UAE President and Secretary General of ATRC, placed the project in national terms. “Our language belongs in the future we are building,” he said. “With Falcon-Emirati, we are putting our knowledge and expertise behind that belief. There is real pride in building technology that understands our people. And there is real freedom in having the capability here in the UAE to develop it further, set our own priorities and turn our ambitions into something people can use.”
Voice and paper carry the Falcon family beyond the chat window
Falcon-ASR converts spoken Emirati Arabic, Modern Standard Arabic, English, French, Spanish and Portuguese into text. The mix reflects the UAE’s population, where Emiratis live and work alongside large expatriate communities. TII said the 1.6-billion-parameter model led on public Arabic speech benchmarks and on internal Emirati speech tests, and that on Emirati speech it beat a 30-billion-parameter multimodal model. The institute did not publish word error rates or name the larger model. Falcon-ASR also time-stamps each word in a recording, which allows accurate subtitles, searchable audio archives and meeting transcripts.
Falcon-OCR-Arabic addresses a different gap. Large volumes of Arabic-language records, from government files to heritage manuscripts, exist only on paper or as scanned images. Arabic script, with its connected letters and right-to-left layout, has historically been harder for character recognition software to read accurately. TII said the model extracts both text and structure, including tables, mathematical formulas and document sections, which suits it to digitisation projects and automated data extraction. No accuracy figures were released.
Dr. Hakim Hacid, Chief Researcher of TII’s Artificial Intelligence and Digital Research Center, said the three models reflected a deliberate choice to specialise. “Across Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic, we have applied targeted specialisation to language, speech and document understanding, addressing areas where Arabic AI still has significant room to advance,” he said.
Dialect models give sovereign AI a test that compute alone cannot pass
The GCC’s sovereign AI strategies have so far been measured largely in data centres, chips and model size. The UAE has backed Falcon and the Jais models developed by G42’s Inception with Mohamed bin Zayed University of Artificial Intelligence, while Saudi Arabia’s SDAIA has built ALLaM for Arabic. Falcon-Emirati adds a narrower measure: whether a model can follow a grandmother’s proverb or a caller’s complaint in the dialect it was spoken in.
The model’s size matters to that test. At 7 billion parameters, Falcon-Emirati is compact enough to run on an organisation’s own infrastructure, and Falcon-ASR is smaller still. If TII makes them available for private deployment, that would suit ministries, banks and telecom operators that must keep citizen data inside national borders. If the benchmark results hold up under independent testing, the approach offers a template for Saudi, Kuwaiti, Omani or Levantine dialect models built on the same Arabic foundation. It would also give public services across the region a route to answer people in the language they use at home.
