Technology Innovation Institute has released Falcon-Emirati-7B, a dialect-specialized large language model built on the Falcon-H1-Arabic architecture to understand and generate Emirati Arabic with native-level accuracy.
The model leverages a hybrid State Space Model and Transformer attention structure to handle long sequences efficiently while maintaining precision for morphologically rich languages. Training involved a dedicated pipeline combining authentic Emirati web data, MSA content focused on local culture, and strictly constrained synthetic data to ensure grammatical and cultural fidelity.
Falcon-Emirati-7B achieves an 84.83% score on the Alyah benchmark, outperforming larger multilingual models. In open-ended generation tests, it demonstrated superior dialect fidelity compared to competitors like ALLaM and Jais, which often defaulted to Modern Standard Arabic even when asked in Emirati.
The release highlights that dialect competence requires targeted training rather than relying on model scale alone, addressing the specific vocabulary, tone, and cultural context of the UAE.