TII's Falcon-Emirati-7B targets Emirati Arabic, but its edge rests on TII's own tests

• Model: Falcon-Emirati-7B, a dialect-specialized model built on Falcon-H1-Arabic-7B (TII blog).
• Purpose: understanding and generating Emirati Arabic, including idioms, proverbs, nabati poetry, etiquette and heritage knowledge.
• Access: the reviewed sources document only the Falcon Chat web interface (TII blog; Arabian Reseller). Neither confirms downloadable weights or a license. The base family was released as open models (Falcon-H1-Arabic blog).

• Architecture: inherited from the base model. Mamba (state-space) and attention layers run in parallel in each block, and their outputs are concatenated before the block's output projection. The 7B base has a 256K-token context (Falcon-H1-Arabic blog).
• Data: three sources (TII blog):
• Emirati web and forum text written natively in dialect
• MSA material about Emirati culture
• Synthetic dialect text, constrained by Emirati glossaries and grammar rules
• Recipe: TII ran ablations on data mix and training stage (continued pre-training, supervised fine-tuning, preference optimization). It does not publish the final mix, token counts or which stage it chose.
• Evaluation: Alyah multiple-choice accuracy, native-speaker review, and an LLM judge ("Gemini 3.7 Flash") scoring open-ended answers for correctness and dialect fidelity, plus blind pairwise comparisons. Prompts, judge rubrics and the synthetic-data generator are not released, so the results cannot be reproduced.

All of the following are vendor-reported. No independent evaluation was found.

• Alyah: 84.83%, ahead of every model TII compared it against. Falcon-H1-Arabic models are excluded from that chart (TII blog). The base 7B-Instruct model scored 82.18% in January (Alyah blog), so the multiple-choice gain over the base is about 2.7 points.
• Dialect fidelity (partial credit): 0.52, against 0.05 for ALLaM, 0.03 for gemma-3-27b-it, 0.02 for Jais-2-8B-Chat and about 0.00 for Fanar-2-27B (TII blog).
• Pairwise judging: wins Poetry 0.88–0.12 against Fanar. Loses Greetings 0.46–0.54 to Jais.
• ArabCulture-Dialogue, UAE subset (283 scenarios, TII's own run): 85.57%, against 83.39% for ALLaM-7B (TII blog). The benchmark itself comes from an outside group (arXiv).

• Falcon-H1-Arabic already claimed the top Open Arabic LLM Leaderboard (OALL) results, with 71.7% for the 7B model (Falcon-H1-Arabic blog; TII news).
• Independent research finds that models handle Arabic dialects worse than MSA:
• ArabCulture-Dialogue (arXiv)
• DialectalArabicMMLU (arXiv)
• AL-QASIDA (arXiv)
• These studies support the problem TII describes. None of them tests Falcon-Emirati.

• Evaluation independence: Alyah was built largely by the same TII authors who built the model (Alyah blog). TII also ran every comparison itself.
• Judge validity: a single proprietary LLM judge scored the open-ended results, and no agreement with human raters is reported.
• Omitted baseline: leaving out the base model makes the gain from specialization hard to see. On multiple choice it appears modest; the large gap is in dialect fidelity.
• TII's own caveats: the model can reflect training-data biases, makes errors on rare or highly local expressions, and native speakers sometimes disagree on the right answer. TII advises testing before any sensitive or official use (TII blog).

• Dialect is a separate requirement. For UAE consumer-facing chat, customer service or government services, strong Arabic benchmark scores do not guarantee replies in the dialect. Teams should measure the register of replies separately from accuracy.
• A replicable pattern. TII's documented approach (native web text, culture-focused MSA, glossary-constrained synthetic data, a native-speaker benchmark) could be applied to other low-resource dialects at 7B scale.
• Procurement. The sources describe chat-only access, so self-hosting needs a confirmed weights release. TII presents the launch as "sovereign AI" capability (Arabian Reseller).