Falcon-Emirati-7B: 84.83% Alyah score tops Arabic AI rivals

Falcon-Emirati-7B did not start from scratch. It builds on Falcon-H1-Arabic, which the Technology Innovation Institute says already set new benchmarks for Arabic earlier this year and was trained on a mix of Modern Standard Arabic and Gulf, Levantine, Egyptian and Maghrebi dialects. That foundation gave the team a model that understood Arabic broadly broadly, handled long context, and had some dialectal exposure built in — but not the Emirati-specific vocabulary, grammar and cultural knowledge.

The team describes dialect adaptation as genuinely hard for three reasons: Emirati is mostly spoken and shows up far less in online writing than Modern Standard Arabic; meaning is often non-literal, leaning on idioms, proverbs and poetry; and there is no established recipe for how much dialectal data to use or which training stage matters most. To work around that, they built a dedicated Emirati data pipeline from three sources — natively written Emirati web content, Modern Standard Arabic material about Emirati culture and heritage, and synthetic dialect data constrained by Emirati glossaries and style rules.

The results suggest the approach worked, but the comparison set is important. All competing models in the study were instruction-tuned Arabic or multilingual models, and Falcon-H1-Arabic family models were excluded because Falcon-Emirati-7B is built on top of them. The one category where competitors hold their own is Greetings & Daily Expressions, which the team notes is where Emirati and Modern Standard Arabic overlap most. Whether dialect-specialized models like this one become the default for regional applications — rather than a niche layer on top of larger general models — likely hinges on whether the dialect-fidelity advantage holds up in real, open-ended use.