Skip to content
by InfinyAI ES

IA en un minutoNewsModels

Models

TII unveils Falcon-Emirati, a language model focused on the Emirati dialect and culture

By

In 30 seconds

The Technology Innovation Institute (TII) has introduced Falcon-Emirati-7B, a language model with 7 billion parameters specialized in the Arabic spoken in the United Arab Emirates. On Alyah, a test of 1,173 questions collected from native speakers, it scores 84.83%, according to TII's own results. It can be tried now in Falcon Chat, though TII warns it can get rare expressions wrong.

The Technology Innovation Institute (TII), the creator of the Falcon models, has introduced Falcon-Emirati-7B, a language model specialized in the Arabic spoken in the United Arab Emirates.

The Arabic of news and textbooks, Modern Standard Arabic, is not the one people use in everyday conversation. In the Emirates people speak a Gulf dialect with its own vocabulary, and its proverbs and poetry do not survive a word-for-word translation.

Falcon-Emirati-7B, with 7 billion parameters, is built on Falcon-H1-Arabic, the institute's own family of Arabic models. It was trained on text from websites and forums written in the Emirati dialect, on material about the country's culture and on synthetic data generated under glossaries and style rules for the dialect.

On Alyah, a test of 1,173 questions collected from native speakers, it scores 84.83%, ahead of every model it was compared against, some of them much larger. In a second test, with a Google model (Gemini 3.7 Flash) acting as judge, it was the only one of the five models compared that reliably answered in Emirati: the others often knew the answer but gave it in Modern Standard Arabic. These are TII's own results.

It can be tried now in Falcon Chat, the institute's chat website. TII warns that it can get rare expressions wrong and recommends testing it before using it for anything official or sensitive.

TII sums it up this way: model size alone is not enough to master a dialect, it has to be trained for on purpose.

Why it matters

TII shows that model size alone is not enough to master a dialect: a small model trained on purpose beats much larger ones.

Official source: TII

Is your website up to scratch? We will audit it for free

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the official source for every one.

Choose one or both:

You will get an email to confirm your subscription: until you click its link, you will not receive anything. You can unsubscribe from any email.