Falcon-Emirati-7B Tops Emirati Arabic and Culture Benchmarks
Falcon-Emirati-7B leads Emirati Arabic benchmarks with 84.83% Alyah accuracy and stronger dialect fidelity, while highlighting sparse data and bias limits.
Summary
On October 6, 2026, the Technology Innovation Institute introduced Falcon-Emirati-7B, a dialect specialist built on Falcon-H1-Arabic. Its hybrid blocks run Mamba state space models and Transformer attention in parallel. The base family spans 3B, 7B and 34B parameters, supports context windows up to 128K and 256K tokens, and covers Modern Standard Arabic, Gulf, Levantine, Egyptian and Maghrebi dialects, English and multilingual data. TII chose 7B because 3B lacked sufficient capacity while 34B imposed unjustified training and serving costs.
Training combined authentic Emirati websites and forums, MSA material about UAE culture and identity, and synthetic dialect data constrained by Emirati glossaries, dictionaries and style rules. With no established adaptation recipe, TII tested data volumes, training stages and authentic to synthetic mixes using automatic scoring and native speaker reviews. Alyah, its 1,173 sample native benchmark, covers greetings, etiquette, figurative language, heritage and poetry.
Falcon-Emirati-7B reached 84.83% Alyah accuracy, leading every compared Arabic and multilingual model. On open answers judged by Gemini 3.7 Flash, its dialect fidelity scored 0.52, versus ALLaM-7B-Instruct-preview at 0.05, gemma-3-27b-it at 0.03, Jais-2-8B-Chat at 0.02 and Fanar-2-27B-Instruct near 0.00. Fanar also scored 0.27 for correctness and abstained 26.2%, versus under 5% for others. Pairwise judging found Falcon strongest in poetry, language, heritage and sensitivity, but it lost greetings to Jais, 0.46 to 0.54, and tied ALLaM at 0.50. On 283 UAE ArabCulture-Dialogue scenarios, Falcon scored 85.57%, ahead of ALLaM at 83.39%, Jais at 73.79% and Fanar at 71.50%. Falcon Chat now offers the model, but rare expressions, localized references, bias and subjective cultural judgments remain limitations.
Positives
- 84.83% Alyah accuracy placed Falcon-Emirati-7B ahead of every Arabic and multilingual model included in the comparison.
- 0.52 dialect fidelity exceeded ALLaM's 0.05, gemma's 0.03, Jais's 0.02 and Fanar's near-zero score.
- 85.57% accuracy across 283 UAE ArabCulture-Dialogue scenarios led ALLaM, Jais and Fanar.
- Native Emirati reviews assessed naturalness, tone and cultural appropriateness beyond automatic benchmark scores.
- Falcon Chat now gives users direct access to Falcon-Emirati-7B for testing.
Risks & concerns
- Written Emirati data remains scarcer than Modern Standard Arabic and several other regional dialects.
- Rare expressions, highly localized references and poorly represented edge cases can still produce incorrect answers.
- Training data may introduce bias, while native speakers can disagree over culturally appropriate wording.
- Gemini 3.7 Flash, rather than native speakers, judged the open-ended and pairwise model comparisons.
- 0.46 versus 0.54 left Falcon behind Jais on Greetings and Daily Expressions.