Falcon-Emirati-7B was released by the UAE's Technology Innovation Institute (TII) through its official Hugging Face blog on October 6, 2026. It is a 7B model built specifically for the Emirati dialect, on top of TII's own Falcon-H1-Arabic, which uses a hybrid architecture that runs Mamba state-space models and Transformer attention in parallel. Its goal is narrow and explicit: to understand and use Emirati Arabic the way a native speaker does, rather than merely speaking Modern Standard Arabic fluently. The model is now available to try on the Falcon chat platform.
Answering correctly in Standard Arabic is not the same as understanding the dialect
Arabic is really a family of languages under one name. News and textbooks use Modern Standard Arabic, but everyday conversation, humor, negotiation and storytelling in the UAE happen in a Gulf dialect. Nabati poetry, proverbs and allusions carry meaning that depends on shared cultural context; a word-for-word translation can be entirely correct and still miss the point. The most common failure of general Arabic models is hearing the question perfectly and then answering in Standard Arabic by default — right content, wrong register. Falcon-Emirati-7B exists to close exactly that gap.
Data from three sources, checked by native ears
The first problem in dialect adaptation is data: Emirati is mostly spoken, so far less native text exists online than for Standard Arabic. TII took three routes at once: crawling and curating authentic Emirati-dialect text from local websites and forums; adding Standard-Arabic material about Emirati culture, history and customs so the model knows what it is talking about; and generating synthetic dialect data constrained by purpose-built glossaries and style rules, so it does not produce sentences that are grammatically fine but instantly sound fake to locals. There was no standard recipe for the training strategy either, so the team ran ablations on data mixes, injection stages and supervision, checking every step with both automatic scores and native-speaker judgment. The 7B size, rather than 34B, was chosen as the balance point between dialect nuance and practical training and serving cost.
Two scorecards: one for being right, one for sounding right
The first is Alyah, a benchmark of 1,173 multiple-choice questions collected manually from native speakers, spanning greetings, figurative language, heritage knowledge and poetry. Falcon-Emirati-7B scores 84.83%, ahead of every other Arabic and multilingual model in the comparison. On the UAE portion of ArabCulture-Dialogue, 283 scenarios, it leads at 85.57%, ahead of ALLaM-7B at 83.39%, Jais-2-8B at 73.79%, and the much larger Fanar-2-27B at 71.50%.
The second scorecard is the more telling one. In open-ended generation, with Gemini 3.7 Flash as judge, models were scored separately on whether they actually answered in the Emirati dialect. Falcon-Emirati-7B reaches 0.52 on dialect fidelity, while ALLaM, gemma-3-27b-it, Jais-2-8B and Fanar-2-27B manage only 0.05, 0.03, 0.02 and effectively zero — the others often know the right answer and say it in Standard Arabic anyway. It does not win everything: in greetings and daily expressions it narrowly loses to Jais-2-8B, the category where dialect and standard overlap most. TII also cautions that rare expressions and highly localized references can still go wrong, and advises evaluating the model before sensitive or official use.
The reusable lesson for multilingual models is the standard it sets: parameter scale does not buy dialect competence, and a long language list on a spec sheet proves nothing. What works is preparing native data and native benchmarks for one dialect at a time, then letting native speakers do the final acceptance — a method that applies well beyond Emirati Arabic.
