84.83%: TII's Emirati model and the race for local AI
"Our language belongs in the future we are building. 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," said H.E. Faisal Al Bannai, Adviser to the UAE President and Secretary General of Abu Dhabi's Advanced Technology Research Council (ATRC).
The Technology Innovation Institute (TII), ATRC's applied research arm, has launched Falcon-Emirati. It is a 7-billion-parameter language model built for Emirati Arabic, including its vocabulary, idioms, expressions and cultural references. The model scored 84.83% on Alyah, a native Emirati Arabic benchmark that covers everyday language, figurative expressions, heritage knowledge and poetry. That puts it ahead of every Arabic and multilingual open-source model TII evaluated. TII released two companion models at the same time. Falcon-ASR is a 1.6-billion-parameter speech recognition model that transcribes six languages. Falcon-OCR-Arabic is a lightweight model that reads Arabic text and structured content from images and documents.
The three models give developers, researchers and everyday users an AI that can work in Emirati Arabic across text, speech and documents. TII sees uses ranging from multilingual customer service and meeting transcription to subtitling, accessibility tools and voice-enabled public services.
Everyday speech is the next test for AI
"Sovereign AI capability must reflect the language used in daily life. 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," said Dr Najwa Aaraj, Chief Executive Officer of TII.
Arabic is spoken by hundreds of millions of people, and it lives in many spoken forms. Modern Standard Arabic is the language of formal writing, education and news. Daily conversation runs through regional dialects, each with its own sound and sense of humour. Emirati Arabic carries much of its meaning in poetry, proverbs and anecdotes, and a model trained mainly on the formal language can understand every word of an Emirati sentence while still missing what the speaker means. Falcon-Emirati is designed to catch that meaning.
To build it, TII researchers started from the institute's Falcon-H1-Arabic model. They added native Emirati content, Modern Standard Arabic material on Emirati culture and heritage, and synthetic data shaped by Emirati-specific linguistic resources. The result is a model that a resident can talk to in the language of their family and neighbourhood.
Compact models, specialist skills
At 1.6 billion parameters, Falcon-ASR is small by current standards. TII reports that it delivered leading accuracy on public Arabic and internal Emirati speech benchmarks, and that it outperformed a 30-billion-parameter multimodal model on Emirati speech. It transcribes Emirati Arabic, Modern Standard Arabic, English, French, Spanish and Portuguese. Those are the languages of a country where Emiratis live and work alongside residents from around the world.
Falcon-ASR also records the moment each word is spoken in an audio file. That makes precise subtitles, searchable recordings and synchronised transcripts possible, which will help viewers who are deaf or hard of hearing as well as journalists, students and anyone working through hours of recorded meetings. Falcon-OCR-Arabic does the same kind of specialist work for documents. It extracts tables, mathematical formulas and document sections from Arabic pages, which supports the digitisation of archives, searchable records and automated information extraction.
"These capabilities empower developers, researchers and communities across the region to build on Emirati language and culture, while expanding what is possible for Arabic AI more broadly. 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," said Dr Hakim Hacid, Chief Researcher of TII's Artificial Intelligence and Digital Research Centre.
A worldwide push for local-language AI
Switzerland made a similar move in September 2025. That month, EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus, an open model trained on more than 1,000 languages, with 40% of its data in languages other than English. Swiss German and Romansh are part of that mix. The model comes in 8-billion and 70-billion-parameter versions and is available through Hugging Face and Swisscom's sovereign AI platform.
India is following a path that closely matches TII's three-model approach. Under the ₹10,300-crore IndiaAI Mission, Sarvam AI has released Sarvam Vision, a 3-billion-parameter document intelligence model for 22 Indian languages that scored 84.3% on olmOCR-Bench. It has also released Bulbul V3, a text-to-speech model with more than 35 voices across 11 Indian languages, built to handle code-switching and regional accents. From Zurich to Bengaluru to Abu Dhabi, national AI programmes are investing in focused models for speech, documents and dialect, the parts of language people use most often in daily life.
Alyah gives dialect AI a common yardstick
TII published Alyah in January 2026 as one of the first benchmarks built specifically for Emirati Arabic. It contains 1,173 multiple-choice items sourced from native speakers, covering greetings, idioms, folklore, poetry and cultural references. In its first round, 54 language models were evaluated. TII's Falcon-H1-Arabic 7B instruct model led at 82.18%, ahead of HUMAIN's ALLaM-7B at 77.24% and Google's Gemma 3 27B at 74.68%. Falcon-Emirati's 84.83% extends that lead. The benchmark is published on Hugging Face, so researchers anywhere can test their own models against the same standard.
TII says Falcon-Emirati will be available through a Falcon Chat platform, which puts the model directly in the hands of everyday users. Earlier Falcon models, including Falcon-H1-34B-Base, are already on Hugging Face under TII's account, so businesses, universities and public services have a familiar route for building on the family. For an Emirati grandmother sharing a proverb, a student transcribing an interview, or a resident calling a public service in their own dialect, the change is a simple one: an AI that understands them the way they speak.
