Abu Dhabi's TII Releases Falcon-Emirati-7B, a Language Model Tuned to Emirati Arabic

Abu Dhabi’s Technology Innovation Institute, the research arm of the Advanced Technology Research Council, said Tuesday it has released Falcon-Emirati-7B, a language model tuned specifically for Emirati Arabic dialect and culture.

The launch matters because Arabic is not one uniform spoken language. Modern Standard Arabic is used in formal writing and media, but everyday speech differs sharply by country and even by community. TII said its new model is designed to handle not just Arabic broadly, but the vocabulary, tone and cultural references specific to the United Arab Emirates. “That’s the gap Falcon-Emirati-7B is built to close,” the institute’s authors wrote in a Hugging Face blog post published Oct. 6 under TII’s tiiuae account.

TII said Falcon-Emirati-7B scored 84.83% accuracy on Alyah, an Emirati benchmark the institute said it released with community input. TII described Alyah as a 1,173-question multiple-choice test built from material manually collected from native Emirati speakers, covering greetings, etiquette, figurative language, poetry and heritage knowledge. “Falcon-Emirati-7B scores 84.83% on Alyah, ahead of every other Arabic and multilingual model we compared it against, including several models many times its size,” the authors wrote.

The institute also reported an open-ended generation test using the same 1,173 prompts and Gemini 3.7 Flash as an LLM judge to score correctness and dialect fidelity. In that test, TII said Falcon-Emirati-7B received a dialect-fidelity score of 0.52, while ALLaM, Gemma, Jais and Fanar scored 0.05 or lower. “Falcon-Emirati-7B is the only one of the five that reliably answers back in the dialect it was asked in,” the authors wrote. On the UAE portion of ArabCulture-Dialogue, a cultural-understanding benchmark, TII said the model scored 85.57% across 283 multiple-choice scenarios, the highest among the four models it tested there. TII said the model is available through its Falcon Chat platform at chat.falconllm.tii.ae.

According to the blog post, Falcon-Emirati-7B is built on Falcon-H1-Arabic, part of TII’s Falcon family. TII said Falcon-H1-Arabic uses a hybrid architecture that combines Mamba, a state-space model approach, with Transformer attention, the mechanism widely used in modern language models. The institute said it chose a 7 billion-parameter version as the best balance between capability and cost for a dialect-focused chat system.

TII said it adapted the model using three types of data: curated web content written in authentic Emirati dialect, including forums and local sites; Modern Standard Arabic material about Emirati culture and identity; and synthetic Emirati-dialect data generated under glossary and style-rule guardrails. In plain terms, that means the model was trained not only on how Emiratis write and speak, but also on background cultural material and machine-generated examples meant to reinforce local usage.

That localization is notable in Arabic AI because strong performance in Modern Standard Arabic does not necessarily translate into strong performance in dialect. A model may answer correctly in formal Arabic yet miss the phrasing, idioms or social cues that matter in everyday conversation. TII’s release is part of a broader regional push to build Arabic-language AI systems, alongside UAE-linked Jais, Saudi-backed ALLaM and Qatar’s Fanar.

There are important caveats. The benchmark results and comparison scores were reported by TII itself, not by an independent third party. TII also acknowledged that the model can reflect biases in its training data, may miss rare expressions or highly localized references, and that dialect and cultural nuance can be subjective. And while the blog post points users to Falcon Chat, the research did not identify a clearly posted standalone model card or downloadable public weights release under the exact name Falcon-Emirati-7B.