TII Releases Falcon-Emirati-7B: Emirati Dialect Adaptation, 84.83% on Alyah Benchmark
According to TII's official blog published on October 6, 2026, Falcon-Emirati-7B is a specialized Emirati-dialect adaptation built on Falcon-H1-Arabic, scoring 84.83% on TII's own Alyah benchmark and clearly leading several competitors in an LLM-judged dialect fidelity evaluation; the model is available for trial on TII's chat platform.
What changed
According to TII's official blog published on October 6, 2026, TII introduced Falcon-Emirati-7B, an Emirati-dialect-specialized adaptation built on Falcon-H1-Arabic. The base family uses the Falcon-H1 hybrid architecture, in which Mamba state-space layers and Transformer attention run in parallel within each module and are fused before projection; the family spans three sizes (3B, 7B, 34B), with context windows up to 128K to 256K tokens, and its training data already includes a mix of Modern Standard Arabic and Gulf, Levantine, Egyptian, and Maghrebi dialects. TII says it chose 7B because it best balances the nuance needed for dialect adaptation against training and inference cost. The data consists of three sources: natively written dialectal web text from Emirati websites and forums, MSA corpora about Emirati culture and identity, and synthetic dialect data constrained by vocabulary lists and style rules. The team says there is no established dialect-adaptation recipe, so it ran ablations on data mix, injection stage, and synthetic-data ratio, while using both automated scoring and native-speaker review. On evaluation, the model scores 84.83% on Alyah (1,173 native multiple-choice questions); TII says it leads all Arabic and multilingual models it compared against, and the Falcon-H1-Arabic family was excluded from that comparison because it is the base. Separately, on the same 1,173 open-ended generation questions with Gemini 3.7 Flash as judge, Falcon-Emirati-7B scores 0.52 on dialect fidelity (partial credit), well above ALLaM-7B's 0.05, gemma-3-27b-it's 0.03, Jais-2-8B-Chat's 0.02, and Fanar-2-27B-Instruct's near 0.00; on the UAE portion of the ArabCulture-Dialogue multiple-choice set it scores 85.57%. The model is available for trial on TII's chat platform.
What it means for developers
The direct impact for developers is that this release makes concrete the question of whether general-purpose Arabic models can handle dialect scenarios directly. TII's evaluations indicate that several competitors may not be wrong in content but default to MSA output even when the user asks in Emirati dialect, which explains why in multilingual customer service, localized marketing, community moderation, or culturally related Q&A, the failure of general Arabic models is often not "answering incorrectly" but "using the wrong register." The 7B scale and hybrid architecture mean single-GPU or small multi-GPU deployment is theoretically feasible, which is practically relevant for teams that need to handle user conversations locally or in private environments and have constraints on data leaving their systems. A distinction is needed: TII's comparison conclusions come from its own Alyah benchmark and its own chosen competitor set, so they reflect a vendor's evaluation stance; the base family being excluded from comparison also means the 84.83% should not be directly aligned with Falcon-H1-Arabic's historical results. Whether dialect ability transfers to other Arabic dialects such as Levantine or Egyptian is not evidenced in the source and should not be assumed. The cultural adaptation evaluation was likewise done only on 283 self-defined UAE scenarios.
Integration and evaluation
Recommendation 1 (data and register constraints): the clearly reported change here is that training data is a three-way mix of native dialectal web text, cultural MSA corpora, and vocabulary-constrained synthetic data, with the synthetic portion limited by vocabulary lists and style rules. If your application has a hard requirement that answers be written in Emirati dialect, the approach is to first sample your own business corpus, build an open-ended generation test set with dialect questions, then split evaluation into two independent dimensions: content correctness and whether the answer is in dialect; the precondition is that you have native speakers available or can afford human review, otherwise you can only rely on automated scoring and cannot capture tone and appropriateness. Measurable criteria are the pass rate of dialect fidelity on your own samples and the rate of switching to MSA, not a single accuracy number. Recommendation 2 (scale and deployment cost): the clearly reported change is that TII did only the 7B tier and publicly explained why, and did not release a 34B dialect version. If your scenario is a high-concurrency, low-latency conversational service, first validate on 7B whether it reaches acceptable naturalness; if your scenario is content generation such as poetry, proverbs, or heritage knowledge, which the source notes are exactly where competitors are weakest but where the model itself may also be affected by low-resource samples, then assess whether it is worth switching to a larger base and adapting it, because the 34B route is not offered in this release. The precondition is that your inference and latency budgets are quantifiable. Measurable criteria are latency and cost per thousand requests at a fixed QPS, and native-speaker pass rates on difficult categories. Recommendation 3 (evaluation comparability): the clearly reported change is that TII used its own Alyah benchmark and excluded the base family from comparison. If you already have an Arabic evaluation process, the approach is to keep your own benchmark as the primary measure, treat Alyah as a reference rather than a replacement, and record competitor versions and judge-model versions. The precondition is that you can reproduce the questions or at least fix the sample set. Measurable criteria are whether conclusions agree between your own benchmark and Alyah, and whether rankings remain stable when the judge model is changed.
Limits and open questions
To be clear: Falcon-Emirati-7B does not currently appear in KimiSeek's live model catalog; the availability described here refers only to TII's own chat platform and does not mean this site has launched the model. The numbers in the source, such as 84.83%, dialect fidelity 0.52, and ArabCulture-Dialogue 85.57%, all come from TII's own or self-selected evaluations (Alyah, Gemini 3.7 Flash as LLM judge, competitors chosen by TII), reflecting a vendor's stance and not independently reproduced by a third party; the base family's exclusion from comparison limits direct comparison with Falcon-H1-Arabic. TII itself also notes in its responsibility and limitations section that the model reflects biases in training data, may err on rare expressions, highly localized references, and edge cases with sparse data, and that dialect and cultural appropriateness are subjective in some scenarios and may not be consistent even among native speakers; it recommends scenario-specific evaluation before sensitive, official, or high-risk uses. The source does not disclose weight license terms, parameter activation method, inference memory and throughput, API pricing and commercial terms, nor evidence of transfer to other Arabic dialects; none of these automatically hold because of an "open trial." The decision recommendations in this article are conditional analysis, not conclusions reported by the source.
Illustrations from the original report
