Abu Dhabi’s Technology Innovation Institute has released Falcon-Emirati-7B, a 7-billion-parameter model tuned for Emirati Arabic rather than Modern Standard Arabic. It is built on the institute’s Falcon-H1-Arabic, a hybrid of Mamba and transformer layers, with a context window of up to 128,000 tokens.
The training mix has three parts: authentic Emirati web content, Modern Standard Arabic cultural material, and synthetic data generated under dialect-specific glossaries and style rules. On Alyah, a set of 1,173 Emirati-dialect multiple-choice questions, TII reports 84.83% accuracy, and 85.57% on the UAE subset of ArabCulture-Dialogue. The figure that shows what the tuning actually bought is open-ended generation, where TII reports a dialect fidelity score of 0.52 against 0.05 for the models it compares against.
Most Arabic language models are trained predominantly on Modern Standard Arabic, which is the language of broadcast and officialdom rather than of anyone’s kitchen. A 7B model is small enough to run cheaply, so the question worth watching is whether dialect-specific tuning at this size holds up against much larger multilingual models on what residents actually ask.
TII says the model can reflect biases in its training data and will sometimes get things wrong, particularly on rare expressions and highly localised references, and advises evaluating it before relying on it for anything sensitive or official.
Source: Technology Innovation Institute, Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance.