The Technology Innovation Institute (TII) in Abu Dhabi has released Falcon-Emirati-7B, a language model built specifically to understand and generate Emirati Arabic — not just Modern Standard Arabic, but the Gulf dialect with its own vocabulary, rhythm, and cultural references that MSA-trained models consistently miss.
Built on top of Falcon-H1-Arabic, the 7-billion-parameter model uses a hybrid architecture combining State Space Models (Mamba) with Transformer attention running in parallel inside every block. The team curated three complementary data streams: authentic Emirati-dialect web content crawled from local forums and websites, MSA material about Emirati culture and heritage, and synthetic data constrained by glossaries and style rules built specifically for the dialect. That last component proved critical — without strict guardrails, synthetic output read as grammatically correct but tone-deaf to native speakers.
The results are striking. On Alyah, a native benchmark of 1,173 samples covering greetings, poetry, heritage knowledge, and figurative language, Falcon-Emirati-7B scores 84.83%, outperforming models many times its size. On dialect fidelity — whether the model actually responds in Emirati rather than defaulting to MSA — it scores 0.52 partial credit versus 0.05 or below for every competitor tested. That gap is close to two orders of magnitude at the low end.
The model is available now on TII's Falcon chat platform. The team is also releasing the Alyah benchmark and associated datasets for the broader research community, hoping to establish a standard for evaluating dialect competence in Arabic language models.
