TII Introduces an AI Model Built to Answer in Emirati Arabic

An Arabic chatbot can know the answer and still reply in the wrong dialect. Technology Innovation Institute introduced Falcon-Emirati-7B on October 6, 2026, targeting Emirati vocabulary, tone and cultural context. The team’s evaluations distinguish two abilities: recognizing the right answer and producing a reply that actually uses Emirati Arabic rather than formal Arabic.

Training beyond formal Arabic

The model builds on the seven-billion-parameter version of Falcon-H1-Arabic, rather than starting from scratch. TII says Emirati adaptation is difficult because the dialect is mostly spoken and has less online text than Modern Standard Arabic, the formal variety used in news and textbooks. Proverbs and poetry also depend on cultural meaning that literal readings can miss.

Three sources for dialect and context

• Native dialect text: Curated Emirati websites and forums supplied everyday expressions and examples of how writers move between Emirati and formal Arabic.
• Cultural material: Modern Standard Arabic articles and references covered Emirati heritage, customs, history and social norms. Their role was to teach context, not dialect style.
• Synthetic examples: Generated dialect data filled topic gaps, with Emirati glossaries, dictionaries and strict style rules constraining the output instead of allowing loosely Gulf-sounding Arabic.

The team experimented with data proportions and training stages, including how to balance authentic and synthetic examples. Native Emirati speakers reviewed outputs for naturalness, tone and cultural appropriateness alongside automatic scoring.

84.83%Alyah multiple-choice accuracy

TII reports this score on Alyah, a 1,173-question benchmark collected manually from native Emirati speakers.

Correct answers, different dialects

Alyah covers greetings, etiquette, figurative language, heritage and poetry. TII says Falcon-Emirati-7B led the Arabic and multilingual models in its comparison, including some much larger systems. That chart excludes the Falcon-H1-Arabic family on which the new model is based.

But choosing among four answers does not test whether a model can generate dialect naturally. For its open-ended evaluation, TII reused the same 1,173 questions and had Gemini 3.7 Flash judge responses from Falcon-Emirati-7B and four competitors: ALLaM-7B-Instruct-preview, gemma-3-27b-it, Jais-2-8B-Chat and Fanar-2-27B-Instruct.

The judge assessed content correctness separately from whether an answer used Emirati Arabic rather than Modern Standard Arabic. On the latter measure, Falcon-Emirati-7B received a partial-credit score of 0.52. ALLaM scored 0.05, Gemma 0.03, Jais 0.02 and Fanar effectively 0.00. These are model-judged scores from TII’s evaluation, not independent measurements.

TII’s interpretation is that competing models often knew the answer but defaulted to formal Arabic, even when prompted in Emirati. Its category breakdown showed the dialect gap across greetings, poetry and other topics, rather than only a narrow set of questions.

For readers considering using the model, the announcement includes a “Try Falcon-Emirati-7B” link. It does not specify a license or distribution format, leaving those practical terms unresolved despite the detailed evaluation results.