The Pulse
TII Releases Falcon-Emirati-7B for Emirati Arabic
The Technology Innovation Institute has released Falcon-Emirati-7B, a language model adapted from Falcon-H1-Arabic to understand and generate Emirati Arabic. TII says it scored 84.83% on its Alyah benchmark and is available through Falcon C
AI.info Team · Written with AI ·
“That's the gap Falcon-Emirati-7B is built to close.”
Shaikha Alsuwaidi, Omar Alkaabi and co-authors, Technology Innovation Institute
A Falcon model trained for Emirati speech
The Technology Innovation Institute (TII) released Falcon-Emirati-7B on October 6, presenting it as a language model specialized in the Emirati dialect. The 7-billion-parameter model builds on Falcon-H1-Arabic and is designed to understand and generate Emirati Arabic, including its vocabulary, grammar and cultural references. TII makes it available through its Falcon Chat platform.
The release addresses a gap between written Modern Standard Arabic (MSA) and everyday conversation in the UAE. TII says Emirati is underrepresented in online writing compared with MSA, while proverbs, poetry and idioms can depend on context that a literal translation misses. A model may understand the words in an Emirati sentence, the institute argues, without grasping its intended meaning or responding in the dialect.
Three sources shaped the adaptation
TII says it started from Falcon-H1-Arabic’s broader Arabic capabilities rather than training a new model from scratch. The parent family uses a hybrid architecture that runs Mamba state-space modeling and Transformer attention in parallel, and includes models with 3 billion, 7 billion and 34 billion parameters. The Emirati adaptation uses the 7-billion-parameter version, which TII selected to balance model quality with training and serving costs.
For the adaptation, the team combined curated text from Emirati websites and forums with MSA material about Emirati culture, heritage and social norms. It also generated synthetic dialect data, guided by glossaries and rules for Emirati vocabulary and grammar. TII says it tested different data mixes and training stages, using native-speaker reviews alongside automatic scores to judge whether responses sounded natural and culturally appropriate.
Alyah tests knowledge and dialect fidelity
TII reports that Falcon-Emirati-7B scored 84.83% on Alyah, a multiple-choice benchmark of 1,173 questions collected from native Emirati speakers. Its questions cover subjects including greetings, etiquette, figurative language, heritage and poetry. TII says the model scored higher than the Arabic and multilingual instruction-tuned models in its comparison; the post excludes Falcon-H1-Arabic from that comparison because it is the model’s base family.
The team also tested whether models would answer open-ended Alyah questions in Emirati Arabic rather than defaulting to MSA. In an evaluation judged by Gemini 3.7 Flash, TII gives Falcon-Emirati-7B a dialect-fidelity score of 0.52 partial credit, compared with 0.05 for ALLaM-7B-Instruct-preview, 0.03 for gemma-3-27b-it and 0.02 for Jais-2-8B-Chat. Fanar-2-27B-Instruct scored effectively 0.00 on that measure. TII describes Falcon-Emirati-7B as the only one of the five models that reliably replied in Emirati Arabic.
A separate cultural evaluation tested 283 UAE scenarios in both Emirati Arabic and MSA. TII reports an overall score of 85.57% for Falcon-Emirati-7B, ahead of ALLaM-7B at 83.39%, Jais-2-8B at 73.79% and Fanar-2-27B at 71.50%. The institute cautions that the model can make mistakes, particularly on rare expressions and highly localized references, and that judgments about cultural nuance can differ even among native speakers.
The limits of the benchmark
The release presents benchmark results produced by TII, not an independent assessment of how the model performs across everyday use. Multiple-choice accuracy measures whether the model can select a listed answer; it does not establish that it will always produce natural Emirati speech in an open conversation. The separate generation test and native-speaker reviews address parts of that question, but TII itself advises users to assess the model for their particular needs before relying on it in sensitive or high-stakes settings.
Falcon-Emirati-7B is available to try on TII’s Falcon Chat platform. Its release focuses on one dialect and pairs the model with Alyah, a benchmark intended to measure dialect-specific language and cultural understanding.