Falcon-Emirati-7B Speaks the Dialect, Not Just the Language
TII's new 7B model targets Emirati Arabic — the Gulf dialect that Standard Arabic models consistently miss — with benchmark scores to match.
By The ai-compute Desk
Arabic is not one language. Modern Standard Arabic fills textbooks and news broadcasts, but in the UAE, real conversation — humor, negotiation, poetry, storytelling — happens in Emirati Arabic, a Gulf dialect with its own vocabulary, rhythm, and cultural freight. A model fluent in MSA can translate every word of an Emirati sentence and still miss what it actually means. [1] That is the specific gap Falcon-Emirati-7B is built to close.
Released by the Technology Innovation Institute, Falcon-Emirati-7B sits on top of Falcon-H1-Arabic, the institute's Arabic model family built on a hybrid architecture that pairs State Space Models (Mamba) with Transformer attention running in parallel inside every block, with their outputs fused before each block's projection. [1] That design trades the quadratic cost of pure attention for the linear-time efficiency of Mamba on long sequences, while keeping attention's precision for long-range dependencies — an important property for a morphologically rich language like Arabic.
The base family spans three scales — 3B, 7B, and 34B parameters — with context windows up to 128K and 256K tokens, trained on a broad mix of MSA and dialectal Arabic across Gulf, Levantine, Egyptian, and Maghrebi varieties, alongside English and multilingual data. [1] TII chose the 7B variant for the dialect-specialized model, calling it the sweet spot: large enough to hold cultural nuance, small enough that training and inference stay practical. The 34B would likely push quality further, but the serving cost doesn't make sense for a dialect-specific chat model.
Dialect adaptation turned out to be harder than it might sound. Emirati Arabic appears far less in writing online than MSA, or even other Gulf and Levantine dialects, so raw training data is scarce. [1] Nabati poetry, proverbs, and figurative expressions lean on shared cultural context rather than surface vocabulary — things no amount of MSA exposure teaches automatically. TII ran three data streams in parallel: crawled content from Emirati websites and forums written natively in the dialect; MSA-language material covering Emirati culture, heritage, and social norms; and synthetic data constrained by glossaries and dictionaries built specifically for Emirati vocabulary and grammar.
Because there is no established recipe for MSA-to-dialect adaptation, training strategy became an experimental question in itself. [1] The team ran ablations on data mix ratios, training stages, and supervision approaches, checking results against both automatic scores and native-speaker review — because automatic metrics alone do not capture naturalness, tone, or cultural fit well enough to trust on their own.
Quantitative tracking used Alyah (الياه, meaning 'North Star'), a 1,173-sample benchmark collected manually from native Emirati speakers, covering everyday greetings, figurative language, heritage knowledge, and poetry — the categories where dialect matters most and where generic Arabic models tend to struggle most. [1] Falcon-Emirati-7B scored 84.83% on Alyah, ahead of every other Arabic and multilingual model tested, including several many times its size. On the UAE portion of ArabCulture-Dialogue, a benchmark for cultural understanding in Arabic (Al Kautsar et al., ACL 2026), it scored 85.57%, ahead of ALLaM-7B at 83.39%, Jais-2-8B at 73.79%, and Fanar-2-27B at 71.50%. [1][2]
The starkest result came from open-ended generation. When prompted in Emirati and judged on whether responses actually came back in the dialect rather than defaulting to MSA, Falcon-Emirati-7B scored 0.52 on dialect fidelity against 0.05 for ALLaM, 0.03 for gemma-3-27b-it, 0.02 for Jais-2-8B-Chat, and effectively 0.00 for Fanar-2-27B-Instruct. [1] The other models often knew the correct answer but delivered it in Standard Arabic. Fanar-2-27B-Instruct also declined to answer 26.2% of the time, versus under 5% for every other model in the comparison.
The results point to something the field keeps relearning: scale is not a substitute for specificity. Building a model that speaks a community's language — not the prestige register it gets mapped onto — requires deliberate data work, targeted evaluation, and the patience to treat cultural fluency as an engineering problem in its own right.
Falcon-Emirati-7B is available now on TII's Falcon Chat platform. If the dialect-fidelity gap closed this sharply at 7B, the question worth watching is what happens when the same pipeline runs at 34B — or when other under-resourced spoken dialects get the same treatment.
Size alone doesn't buy you dialect competence. Some of the largest multilingual models here score well below smaller, more dialect-aware ones, which tells you Emirati proficiency has to be trained for on purpose, not picked up as a side effect of scale.TII Falcon team