# Arabic language models for Gulf operations: a 2026 survey

Survey. Qantara Research, Qantara. 14 October 2026. 15 min read.

> Which Arabic text and speech models can carry a Gulf company's orders, messages, voice notes and documents, what the leaderboards measure, and what licences and hosting allow.

## Abstract

Which Arabic language and speech models can an agent rely on when work arrives as a Gulf-dialect WhatsApp message, a voice note or a scanned invoice? We surveyed the Arabic-centric and Arabic-adapted models released between August 2023 and October 2026, reading their technical reports, model cards and licences, and the benchmarks used to rank them, as published on 14 October 2026. Open-weight models built for Arabic in the Gulf now rank alongside multilingual models several times their size on native Arabic benchmarks. Most of those benchmarks, however, test Modern Standard Arabic, some were translated from English or carry wrong answer keys, and a new Emirati benchmark finds models that understand the dialect but answer in the standard language. Speech is the weakest link: on one Saudi television test set, open recognisers get 48% to 93% of Gulf dialect words wrong, against 19% to 66% of standard Arabic words. Licences split the field into weights a company may host in its own country and models reachable only as a service. For Gulf operations, this means choosing a model per workflow, testing it on the company's own past cases, and hosting open weights in-country where data cannot leave.

## Key findings

1. Arabic-centric open models now compete at the top of native Arabic leaderboards: in Qimma's April 2026 ranking, Jais-2-70B-Chat placed third with 65.81, behind Qwen3.5-397B and Karnak [1].
2. Most Arabic benchmarks measure Modern Standard Arabic, and their quality varies: the Open Arabic LLM Leaderboard dropped its machine-translated tasks in 2025, and Qimma discarded 3.1% of ArabicMMLU's questions as flawed [2, 1].
3. Models understand Gulf dialect better than they write it: in TII's judged test, Falcon-Emirati-7B answered in Emirati with a fidelity of 0.52, against 0.05 or less for four other models [3, 4].
4. Speech recognition is the weakest link: on the Sada test set, every system in the Open Universal Arabic ASR Leaderboard made more word errors on Khaliji speech than on Modern Standard Arabic [5].
5. Licence and access differ as much as accuracy: Jais 2, Fanar and AceGPT publish Apache 2.0 weights, Aya and Command R7B Arabic weights are non-commercial, and ALLaM 34B and Falcon-H1-Arabic were reachable only as services when we checked [6, 7, 8, 9, 10, 11].

## 1. Introduction

The work of a Gulf distributor arrives in several kinds of Arabic at once: an order typed in Gulf dialect on WhatsApp with the product names in English, a voice note from a loading bay, a supplier invoice in Modern Standard Arabic and English, a customs declaration in formal Arabic. An agent that takes part in this work has to read all of it, reply in a register the customer accepts, and run where the company's data may be processed. Arabic is a case of diglossia (Two varieties of one language with different roles: in Arabic, Modern Standard Arabic for writing, news and formal speech, and the local dialect for everyday talk.), and Gulf operations happen on both sides of it.

The question is newly answerable. Between December 2025 and October 2026 the Gulf's main laboratories released Jais 2, Falcon-H1-Arabic, Fanar 2.0 and Falcon-Emirati [12, 11, 13, 3], with new public tests of a Gulf dialect: Alyah for written Emirati [14] and Emirati speech results for Falcon-ASR [15]. Arabic remains scarce in training data, about 0.5% of the web by the Fanar team's estimate, despite 400 million native speakers [13], so what a model was trained on, and in which variety, matters as much as its size.

Existing surveys describe Arabic models and benchmarks well [16, 17]. This one asks a company's question: which models can be trusted with which kinds of Gulf work, how far the published scores hold, and what licence and hosting allow. It contributes:

1. A dated inventory of 15 Arabic-centric and Arabic-adapted text models, August 2023 to October 2026, with organisation, sizes, licence and release date.
2. A reading of the main Arabic benchmarks and their known weaknesses.
3. A summary of the published evidence on Gulf dialect, in text and in speech.
4. What follows for choosing, testing and hosting a model per workflow.

## 2. Background and related work

Al-Khalifa and colleagues trace Arabic language technology from morphological analysers to the large models that followed ChatGPT, among them Jais, ALLaM, AceGPT and Fanar [16]. They note two problems that recur here: because digital Arabic is scarce, many datasets are translated from English and carry its cultural assumptions, and models trained mainly on Modern Standard Arabic struggle with colloquial input. Alzubaidi and colleagues at the Technology Innovation Institute (TII) in Abu Dhabi review more than 40 Arabic benchmarks, keep a public list of them, and name three gaps: little evaluation over time, little of multi-turn dialogue, and cultural misalignment in translated sets [17]. Ouda's list for Silma AI, of March 2025, records more than 25 Arabic-capable models with sizes and licences [8].

Arabic speech recognition grew up on broadcast television. The Multi-Genre Broadcast challenges built MGB-2 from about 1,200 hours of Al Jazeera, mostly Modern Standard Arabic, then Egyptian and Moroccan sets [5, 18]. Gulf speech entered public benchmarks in 2024, with the Saudi Sada corpus [19] and the Emirati part of Casablanca [18]. The Open Universal Arabic ASR Leaderboard of Wang, Alhmoud and Alqurishi compares open recognisers across these sets and is our main speech source [5].

## 3. Scope and method

We included text models that are Arabic-centric (trained from scratch with a large Arabic share) or Arabic-adapted, by continual pre-training (Training an existing model further on new text, here mostly Arabic, instead of training a new model from scratch.) or tuning of a multilingual base, released from August 2023 to 14 October 2026 with a public paper, model card or announcement. Open multilingual models enter where Arabic leaderboards rank them; closed frontier models through the Arabic results their developers publish. For speech, we included public data and evaluations with Gulf content.

Every source was read on 14 October 2026. We say when a number comes from the developer of the model it describes. Release dates are those of the first paper or weights; where a source gives none, we use the creation date of the model's Hugging Face repository. We ran no models. Arabizi, Arabic written in Latin letters, is left to our companion paper How the Gulf writes.

## 4. The model families

**Figure 1. Arabic model releases and Gulf-dialect tests, August 2023 to October 2026** Read down: each entry is a public release, dated by month; the entries aimed at a Gulf dialect date from 2026.

- Aug 2023: Jais 13B. An Arabic-centric model pretrained on Arabic and English, with open weights.
- Sep 2023: AceGPT. Arabic localisation, aligned with a reward model attuned to local culture.
- Oct 2024: Aya Expanse 8B and 32B. Cohere Labs: 23 languages including Arabic, under a non-commercial licence.
- Jan 2025: Fanar Star and Fanar Prime. 7B and 9B models from Qatar, with dialect speech recognition.
- Feb 2025: ALLaM 7B preview and Command R7B Arabic. Open 7B and 8B models from Saudi Arabia's SDAIA and from Cohere Labs.
- May 2025: Falcon-Arabic 7B. TII adapts Falcon 3 with 32,000 added Arabic tokens.
- Aug 2025: ALLaM 34B in Humain Chat. A service hosted in Saudi Arabia, without public weights.
- Dec 2025: Jais 2, 8B and 70B. A custom Arabic vocabulary; dialects and Arabic-English code-switching.
- Jan 2026: Falcon-H1-Arabic and the Alyah benchmark. Models naming Gulf among their dialects, and a benchmark for Emirati.
- Mar 2026: Fanar 2.0, 27B. Continual pre-training of Gemma 3 on 120 billion curated tokens.
- Apr 2026: Qimma leaderboard. Validates benchmark questions before ranking 46 models.
- Oct 2026: Falcon-Emirati 7B and Falcon-ASR. A model specialised in Emirati, and a speech recogniser with Emirati results.

Source: the papers, model cards and announcements cited in Table 1 and in the text [20, 21, 9, 22, 23, 24, 10, 6, 11, 14, 13, 1, 3, 15]. Months without a stated date are Hugging Face repository creation dates, read on 14 October 2026.

### Abu Dhabi: Jais and Falcon

Jais, released in August 2023 by Inception, the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Cerebras, was a 13B model pretrained on Arabic, English and code [20], followed in 2024 by a family from 590M to 70B parameters under Apache 2.0 [8]. Jais 2 (December 2025, 8B and 70B) is, its authors write, to their knowledge the largest open Arabic-centric model trained from scratch, under a commercially permissive licence [12]. Its card lists a custom Arabic-centric vocabulary of 150,272 tokens and coverage of Modern Standard Arabic, regional dialects and Arabic-English code-switching, without naming the dialects [6]. Vocabulary matters for Arabic because of tokenizer fertility (The average number of tokens a tokenizer cuts a word into. A tokenizer built mostly for English splits Arabic words into more pieces, which raises cost and shortens the text a model can hold at once.).

TII began by adapting its own models. Falcon-Arabic (May 2025) added 32,000 Arabic tokens to Falcon 3-7B and continued training on native, not machine-translated, Arabic [24]. Falcon-H1-Arabic (January 2026) builds on the hybrid Mamba-Transformer Falcon-H1 in 3B, 7B and 34B sizes, with contexts up to 256K tokens, trained on about 300 billion tokens; TII names Egyptian, Levantine, Gulf and Maghrebi among its expanded dialect sources [11]. Falcon-Emirati-7B (October 2026) specialises the 7B in Emirati with Emirati web text, Modern Standard Arabic writing on Emirati culture and synthetic Emirati text [3]. TII offers these through its chat service; we found no public weights for them in its Hugging Face organisation on 14 October 2026.

### Saudi Arabia and Qatar: ALLaM and Fanar

ALLaM, from the National Center for AI at the Saudi Data and AI Authority (SDAIA), was presented in July 2024 and at ICLR 2025; its authors show that vocabulary expansion and mixed Arabic and English pretraining teach a model Arabic without catastrophic forgetting of English [25]. A 7B preview was released under Apache 2.0 [8]. ALLaM 34B, launched in August 2025 inside the Humain Chat app, is reported as built from scratch and hosted end to end in the Kingdom, with no mention of public weights [10].

Fanar, from the Qatar Computing Research Institute at Hamad Bin Khalifa University, paired a 7B model trained from scratch on nearly 1 trillion tokens with a 9B model continually trained from Gemma 2, alongside speech recognition for several dialects [22]; the 9B weights are under Apache 2.0 [8]. Fanar 2.0 (March 2026) centres on a 27B model continually pretrained from Gemma 3 on 120 billion curated tokens, built and run entirely at the institute, reporting 3.5 points gained on dialect benchmarks over Fanar 1.0 and a long-form speech recogniser [13]. Its card names Gulf, Levantine and Egyptian among supported dialects, under Apache 2.0 [7].

### Cohere and other Arabic-tuned models

Cohere Labs' Aya programme runs from Aya 101 to Tiny Aya, a 3.35B model for more than 70 languages [26]. Aya Expanse covers 23 languages including Arabic at 8B and 32B, and Aya Vision adds images, both under non-commercial CC-BY-NC 4.0 [9, 8]. Command R7B Arabic (February 2025), an 8B model, is described as optimised for Modern Standard Arabic and English, under the same licence [23]. AceGPT, from the Chinese University of Hong Kong, Shenzhen, with King Abdullah University of Science and Technology, localised an existing model with Arabic pre-training, native Arabic instructions and a reward model attuned to local culture [21]; its 32B second version ranks tenth in Qimma [1]. Silma 9B is built on Google's Gemma under Gemma's terms [8].

### Frontier multilingual models

Open multilingual models compete directly: Qwen3.5-397B leads Qimma with 68.06, and Llama-3.3-70B-Instruct scores 63.96 [1]. Closed models are measured on MMLU translated by professional translators: OpenAI reports 0.904 in Arabic for o3-high, against its 14-language average of 0.888 [27], and Anthropic reports Claude Sonnet 4.5 at 97.2% of its English score in Arabic [28]. These show that frontier models read formal Arabic well. They say nothing about a Gulf voice note or a reply in dialect, and no closed model appears in the Qimma or Alyah tables we read.

**Table 1. Arabic-centric and Arabic-adapted text models, by release date**

| Model | Organisation | Sizes | Licence | Released |
| --- | --- | --- | --- | --- |
| Jais | Inception, Cerebras, Mohamed bin Zayed University of Artificial Intelligence | 13B | Open weights | Aug 2023 |
| AceGPT | Chinese University of Hong Kong, Shenzhen, with partners | 13B; 8B, 32B in v2 | Apache 2.0 | Sep 2023 |
| Jais family | Inception, Cerebras, Mohamed bin Zayed University of Artificial Intelligence | 590M to 70B | Apache 2.0 | Aug 2024 |
| Silma 9B | Silma AI | 9B | Gemma terms | Aug 2024 |
| Aya Expanse | Cohere Labs | 8B, 32B | CC-BY-NC 4.0 | Oct 2024 |
| Fanar Star, Prime | Qatar Computing Research Institute | 7B, 9B | Apache 2.0 (9B) | Jan 2025 |
| ALLaM 7B preview | SDAIA | 7B | Apache 2.0 | Feb 2025 |
| Command R7B Arabic | Cohere Labs | 8B | CC-BY-NC 4.0 | Feb 2025 |
| Aya Vision | Cohere Labs | 8B, 32B | CC-BY-NC 4.0 | Mar 2025 |
| Falcon-Arabic | TII | 7B | No public weights found | May 2025 |
| ALLaM 34B | SDAIA, Humain | 34B | App only | Aug 2025 |
| Jais 2 | Inception, Cerebras, Mohamed bin Zayed University of Artificial Intelligence | 8B, 70B | Apache 2.0 | Dec 2025 |
| Falcon-H1-Arabic | TII | 3B, 7B, 34B | No public weights found | Jan 2026 |
| Fanar 2.0 | Qatar Computing Research Institute | 27B | Apache 2.0 | Mar 2026 |
| Falcon-Emirati | TII | 7B | Chat service only | Oct 2026 |

From the sources cited in the text, read on 14 October 2026 [8, 23]. Dates for the Jais family, Silma 9B, Aya Expanse, ALLaM 7B preview and Aya Vision are Hugging Face repository creation dates; "No public weights found" records our search of TII's Hugging Face organisation that day.

**Figure 2. What each model allows a Gulf company to do** Each row is a model; a filled mark means the property holds, a half mark that it holds in part, and text gives the detail where yes or no would mislead.

|  | Weights published | Licence allows commercial use | Gulf dialect named | Sizes |
| --- | --- | --- | --- | --- |
| Jais 2 | yes | yes | partial | 8B, 70B |
| Fanar 2.0 | yes | yes | yes | 27B |
| ALLaM 7B preview | yes | yes | Not stated | 7B |
| AceGPT v2 | yes | yes | Not stated | 8B, 32B |
| Silma 9B | yes | Gemma terms | Not stated | 9B |
| Aya Expanse | yes | no | Not stated | 8B, 32B |
| Command R7B Arabic | yes | no | no | 8B |
| Falcon-H1-Arabic | None found | Service terms | yes | 3B, 7B, 34B |
| Falcon-Emirati | None found | Service terms | yes | 7B |
| ALLaM 34B | no | Service terms | Not stated | 34B |

Source: Qantara, compiled from the sources in Table 1, read on 14 October 2026. Gulf dialect is yes where the developer names Gulf or a Gulf dialect, partial where it names dialects without saying which, and no where the card names Modern Standard Arabic only.

## 5. Benchmarks and leaderboards

The Open Arabic LLM Leaderboard on Hugging Face is the reference most developers quote. Its second version, of February 2025, removed saturated and machine-translated tasks and added ArabicMMLU, a human-translated MMLU, the AraTrust safety set and a retrieval task scored by a model [2]. ArabicMMLU has 14,575 multiple-choice questions in Modern Standard Arabic from school exams in eight countries; at release GPT-4 scored 72.5% and the best Arabic-centric model 62.3% [29]. Balsam has 78 tasks and 52,000 examples, with blind test sets [30]. AraBench tests translation from dialect into English, at levels down to 25 cities [31]. Qimma, released by TII in April 2026, validates questions before ranking models: two models score each item, native speakers review the flagged ones, and the suite keeps over 52,000 samples, 99% native Arabic [1]. Five weaknesses run through them.

Translation. The leaderboard's maintainers say the tasks translated from English in its first version introduced linguistic and contextual mismatches [2]; the TII survey counts cultural misalignment in translated sets among the field's gaps [17]. Even the dialect benchmark AraDiCE was built by machine translation and human post-editing [32], and every frontier score above is a translation.

Modern Standard Arabic. ArabicMMLU's Gulf sources are 184 questions from the UAE, 111 from Kuwait and 104 from Saudi Arabia, against more than 6,000 from Jordan [29]: by our sum, 399 of 14,575. Alyah's authors note that most Arabic benchmarks focus almost exclusively on the standard language [14].

Wrong answers and broken scoring. Qimma discarded 436 of 14,163 ArabicMMLU items (3.1%) for wrong gold answers, corrupt text and duplicates, and rewrote 88% of the Arabic HumanEval+ prompts [1]. ArabicMMLU's own authors found 96% of a sample of answer keys correct and treat that as the highest meaningful score [29]. The leaderboard found a scoring bug that had cut some small models' scores on one task by up to 20 points [2].

Contamination. Most test sets are static and public, so data contamination (When test questions, or close copies of them, appear in a model's training data, so that a high score partly measures memory rather than skill.) is hard to rule out; Balsam's authors built blind test sets for this reason [30].

Who runs the test, and how. TII scored Falcon-H1-Arabic with a different inference engine from the leaderboard's, a difference it puts at under one point, and scored a newer ALLaM release higher than the leaderboard's [11]; TII also built Alyah and Qimma, which rank its models. None of this suggests bias, but it is a reason to rerun a shortlist. And most of these tests score multiple-choice answers by likelihood [1], while an agent writes replies and fills fields.

## 6. Gulf dialect beside the standard and the larger dialects

Arabic's dialects include Gulf, Levantine, Egyptian and Maghrebi [5], and Gulf is itself several: Sada labels Najdi, Hijazi and Khaliji speech [5], and the Habibi speech work lists Emirati, Omani and Dhofari separately [33]. Benchmarks have favoured the larger dialects: AraDiCE's 45,000 post-edited samples cover Egyptian and Levantine, and only its cultural test spans the Gulf [32]. On AraDiCE, Arabic-specific models such as Jais and AceGPT beat multilingual ones on dialect tasks; models chose Gulf replies best, which the authors attribute to Gulf's closer resemblance to the standard language, and asked to write dialect, they often explained the prompt instead [32].

Alyah, released by TII in January 2026, is built for one Gulf dialect: 1,173 four-option questions collected by hand from native Emirati speakers [14]. Its hardest categories were Language and Dialect and Greetings and Daily Expressions, which the authors attribute to Emirati being mostly spoken and rarely written. Figure 3 shows the instruction-tuned results.

**Figure 3. Accuracy of instruction-tuned models on Alyah, the Emirati-dialect benchmark, in percent** Each bar is one model's accuracy, in percent, on 1,173 four-option questions written by native Emirati speakers; guessing would score 25%.

|  | % |
| --- | --- |
| Falcon-H1-Arabic 7B Instruct | 82.18 |
| ALLaM 7B Instruct preview | 77.24 |
| Gemma 3 27B it | 74.68 |
| Qwen2.5 72B Instruct | 74.6 |
| Aya Expanse 32B | 73.66 |
| AceGPT v2 32B Chat | 72.8 |
| Llama 3.3 70B Instruct | 69.74 |
| Fanar 1 9B Instruct | 69.22 |
| Command R7B Arabic | 64.54 |
| Silma 9B Instruct | 63.94 |
| Aya Expanse 8B | 61.21 |
| Llama 3.1 8B Instruct | 46.29 |
| Qwen3 8B | 25.66 |

Source: TII, Alyah results table for instruction-tuned models, published 27 January 2026 and read on 14 October 2026 [14]; 13 of the 31 models listed, values quoted exactly. TII built both Alyah and the top model; the Jais 2 models are discussed in the post but are not in its table.

Arabic-trained 7B models sit above multilingual models ten times their size: 82.18 for Falcon-H1-Arabic 7B and 77.24 for ALLaM 7B, against 74.6 for Qwen2.5-72B and 69.74 for Llama-3.3-70B [14]. Scores near 25% need care: Qwen3's instruction-tuned models scored 25.66 to 26.26, the level of guessing, and Qimma also singles out Qwen3 as the family whose instruction tuning did not help [1]. We read this as a mismatch between output and scoring code rather than ignorance, and as a reminder of how fragile one score is.

Writing the dialect is harder than reading it. Al-Qasida, across nine models and eight dialects, found that models produce dialect less well than they understand it because they are reluctant to, that post-training can add to the bias, and that a few examples in the prompt overcome it [4]. TII measured the same for Emirati with an LLM-as-judge (Scoring free-text answers by asking another language model to grade them against a rubric. Cheaper than human raters, but it inherits the judging model's biases.): asked the Alyah questions openly, Falcon-Emirati-7B answered in Emirati with a fidelity of 0.52, against 0.05 for ALLaM-7B, 0.03 for Gemma-3-27B, 0.02 for Jais-2-8B and about 0.00 for Fanar-2-27B. The others, the authors write, "often know the right answer but say it in Modern Standard Arabic by default" [3].

The register of a reply is therefore a decision made per workflow: a formal reply may suit a supplier and not a customer who wrote in dialect. Messages also mix languages; Jais 2's card names Arabic-English code-switching (Moving between languages within one conversation or sentence, as in a Gulf message that mixes Arabic with English product names and numbers.) as a target [6].

## 7. Speech: recognising and speaking Gulf Arabic

A voice note is two problems in sequence: turning Gulf speech into text, then understanding the text. If recognition drops a quantity, no language model downstream recovers it. The measure is word error rate (The share of words a speech recogniser gets wrong, by substituting, dropping or inserting them, against a human transcript. Lower is better.), and public data for Gulf speech is small (Table 2).

**Table 2. Public speech data with Gulf content**

| Dataset | Size | Source | Gulf content |
| --- | --- | --- | --- |
| MGB-2 (2016) | About 1,200 hours | Al Jazeera broadcasts [5] | One of five dialect groups; 78% standard Arabic [18] |
| Sada (2024) | 668 hours | 57 Saudi television shows [19] | Mostly Saudi dialects |
| Casablanca (2024) | About 48 hours, 6 Emirati | Television series on YouTube [18] | Emirati among eight dialects |
| Mixat (2024) | 15 hours | Two Emirati podcasts [34] | Emirati mixed with English |
| Habibi test (2026) | 1,375 Saudi and 1,048 Emirati clips | Sada and other corpora [33] | Two of seven speech synthesis subsets |

Figures as stated by each dataset's authors; MGB-2's share of standard Arabic as reported by the Casablanca authors.

The leaderboard's first report ran 14 open system set-ups, without fine-tuning, on five multi-dialect test sets, and its Table 3 splits the Sada test set by dialect [5]. Every set-up did best on Modern Standard Arabic and markedly worse on Egyptian and Khaliji, which the authors attribute to training data dominated by the standard language; Whisper, for one, saw 739 labelled hours of Arabic among 680,000 [5, 35]. Rankings tracked labelled Arabic data, noise and music raised every error rate, and elderly and female speakers fared worse with most systems.

**Figure 4. Word error rate on Modern Standard Arabic and on Khaliji speech in the same Sada test set, in percent** For each system, the first bar is the percentage of words wrong on Modern Standard Arabic and the second on Khaliji (Gulf) speech, both from Saudi television; shorter is better.

|  | Modern Standard Arabic | Khaliji |
| --- | --- | --- |
| Conformer-CTC large with language model | 19.23 | 48.23 |
| Whisper large-v3 | 27.95 | 59.92 |
| Whisper large-v3 turbo | 28.93 | 65.04 |
| Whisper large-v2 | 29.13 | 63.06 |
| SeamlessM4T v2 large | 39.87 | 55.73 |
| MMS 1B all | 42.71 | 83.43 |
| Whisper small | 49.41 | 93.16 |

Source: Wang, Alhmoud and Alqurishi, Open Universal Arabic ASR Leaderboard, Table 3, arXiv, 18 December 2024 [5]; seven of the 14 set-ups, values quoted exactly (the paper prints decimal commas).

Emirati speech tells the same story. On Casablanca's Emirati test set, Whisper large-v3 got 62.31% of words wrong before text normalisation and 52.88% after, and on code-switched segments it produced no words in the other language, transliterating or translating them [18]. Mixat's authors found existing recognisers perform poorly on Emirati podcasts, and worse with code-switching [34].

The Gulf laboratories have begun to answer. Fanar has had dialect speech recognition since its first release, and Fanar 2.0 adds a long-form recogniser [22, 13]. TII's Falcon-ASR (October 2026), a 1.6B model trained on Emirati, other Gulf and Arabic dialects, the standard language and English, reports an average word error rate of 20.92% on the leaderboard's six current test sets, against 23.17% for the best entry it checked on 30 September 2026, and 22.73% on an internal Emirati set [15]. These are the developer's figures, the Emirati set is not public, and the model was offered as a demonstration with an API planned.

For an agent that answers a voice note with a voice note, synthesis matters too. Habibi, released openly in January 2026, trains one model across more than 12 regional dialects and publishes a multi-dialect synthesis benchmark with Saudi and Emirati subsets; its authors report quality competitive with ElevenLabs' Eleven v3 (alpha) [33].

## 8. Open weights, services and in-country hosting

Where a model can run depends on whether its weights are published and what their licence allows. Open weights (A model whose trained parameters can be downloaded and run on one's own hardware. The licence attached to the weights decides whether commercial use is allowed.) under Apache 2.0, as for Jais 2, the ALLaM 7B preview, Fanar and AceGPT, can run in a cloud region in the country, in the company's own cloud or on its hardware, and be used commercially [6, 8, 7]. CC-BY-NC weights, as for Aya and Command R7B Arabic, can be tested, but commercial use falls outside the licence [9, 23].

The rest are services. ALLaM 34B runs inside Humain's app, hosted in Saudi Arabia [10]; TII's Falcon-H1-Arabic, Falcon-Emirati and Falcon-ASR were offered through its chat service and a demonstration [11, 3, 15]; frontier models are reached through their developers' APIs. A service in the same country can meet a residency rule that one abroad cannot, but either way the provider's terms decide where the model runs and what it keeps. Size need not be a barrier either: TII positions its 7B as the general-purpose production model and its 3B for on-device use [11], and Fanar 2.0's authors call sovereignty a first-class design principle [13].

## 9. Implications for deployment

Three conclusions follow for a Gulf company, and they describe how Qantara approaches the work.

Choose per workflow. No model leads on every kind of Arabic work. One that reads invoices well may not reply in dialect, and the best text model cannot recover a quantity the recogniser dropped. Table 3 is an illustrative example of how the choice divides.

**Table 3. An illustrative mapping of Gulf workflows to what to test and where the model can run**

| Workflow | What arrives | What to test on past cases | Where the model can run |
| --- | --- | --- | --- |
| Orders by WhatsApp | Gulf dialect with English product names | Items and quantities; replies in the customer's register | In-country open weights where messages carry personal data |
| Voice notes | Gulf speech on phones, with noise and English words | Word error rate on the company's own notes | In-country recognition where audio cannot leave |
| Supplier documents | Standard Arabic and English, often photographed | Fields extracted against the ERP | Depends on what the document holds |
| Customer replies | Drafts in dialect or standard Arabic | Register and tone, marked by the team | Any model the data allows; a person approves above thresholds |

Illustrative; ours. A method, not measured results.

Evaluate on the customer's own Gulf cases. Section 5 shows why a leaderboard can shortlist a model but not choose it. The test that counts is built from the company's past work: its WhatsApp orders in its customers' dialect, its voice notes from its own warehouses, its supplier documents. At Qantara, each agent is run on the customer's past cases before it goes live, with their team marking what good looks like, and our engineers record a baseline first. Al-Qasida's finding that a few examples restore dialect output suggests the same cases can improve a model as well as measure it [4]. Our technical report Compounding agents: evaluation and A/B testing on live operations sets out the method.

Host where the data must stay. The record of what each agent read and did, and its log, are kept in the customer's country, and for each workflow we state in writing which model reads what and where it runs. Where data cannot leave, open-weight Arabic models from Gulf developers, under licences that allow commercial use, now exist at sizes up to 70B. Where data may leave, frontier models from Silicon Valley, which Qantara brings to Gulf companies through partners, are candidates too. Either way each agent signs in with credentials of its own, every read and write is logged, and above the customer's thresholds a person approves. As Figure 1 shows, the field changes every few months, so a model chosen today should be re-tested on the same past cases when a better one appears.

## 10. Limitations

- We ran no models. Many scores come from the developers of the models they rank, including TII for Alyah, Qimma and the Falcon models.
- Gulf evidence is thin. Alyah covers written Emirati; Sada, Saudi television; Casablanca, six hours of Emirati television; Mixat, 15 hours of podcasts. We found no public benchmark for Omani, Kuwaiti, Bahraini or Qatari text, and none built from business messages or calls.
- Most text benchmarks are multiple choice; they do not test extraction, register or action.
- Closed frontier models are absent from the Arabic leaderboards we read, and their published Arabic figures are translations.
- Speech results come from broadcast and podcast audio; phone voice notes are likely harder, but we found no public measurement.
- Licences, access and availability change; our statements describe 14 October 2026.

A public Gulf-wide benchmark built from operational messages and voice notes would test these conclusions directly; we would welcome one.

## 11. Conclusion

By October 2026, open-weight Arabic models from Abu Dhabi, Riyadh and Doha rank alongside multilingual models several times their size, under licences that let a company run them in its own country. The evidence for Gulf dialect is newer and thinner: models understand it better than they write it, most benchmarks do not test it, and speech recognition, the first step for every voice note, still gets far more Gulf words wrong than standard ones.

For a company putting agents into its operations, published scores are a starting list. The decision belongs to its own past cases, marked by its own team, with each workflow's model chosen, tested and hosted where its data is allowed to be.

## References

[1] AlQadi, L., Alzubaidi, A., Alyafeai, M., et al. (2026). QIMMA قِمّة ⛰: A Quality-First Arabic LLM Leaderboard. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard (accessed 14 October 2026)
[2] El Filali, A., Aloui, M., Husaain, T., et al. (2025). The Open Arabic LLM Leaderboard 2. Hugging Face blog. https://huggingface.co/blog/leaderboard-arabic-v2 (accessed 14 October 2026)
[3] Alsuwaidi, S., Alkaabi, O., Alhammadi, M., et al. (2026). Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/falcon-emirati (accessed 14 October 2026)
[4] Robinson, N. R., Abdelmoneim, S., Marchisio, K., et al. (2025). AL-QASIDA: Analyzing LLM Quality and Accuracy Systematically in Dialectal Arabic. Findings of ACL 2025. https://aclanthology.org/2025.findings-acl.1137/
[5] Wang, Y., Alhmoud, A. and Alqurishi, M. (2024). Open Universal Arabic ASR Leaderboard. arXiv:2412.13788. https://arxiv.org/abs/2412.13788
[6] Inception (2025). Jais-2-70B-Chat model card. Hugging Face. https://huggingface.co/inception42/Jais-2-70B-Chat (accessed 14 October 2026)
[7] Qatar Computing Research Institute (2026). Fanar-2-27B-Instruct model card. Hugging Face. https://huggingface.co/QCRI/Fanar-2-27B-Instruct (accessed 14 October 2026)
[8] Ouda, K. (2025). Arabic LLM Models. Hugging Face blog, SILMA AI. https://huggingface.co/blog/silma-ai/arabic-llm-models-list (accessed 14 October 2026)
[9] Cohere Labs (2024). aya-expanse-8b model card. Hugging Face. https://huggingface.co/CohereLabs/aya-expanse-8b (accessed 14 October 2026)
[10] Middle East AI News (2025). HUMAIN Chat goes live powered by ALLaM 34B LLM. Middle East AI News, 25 August 2025. https://www.middleeastainews.com/p/humain-chat-live-allam-34b-llm (accessed 14 October 2026)
[11] Boussaha, B. E. A., Alyafeai, M., Alzubaidi, A., et al. (2026). Introducing Falcon-H1-Arabic: Pushing the Boundaries of Arabic Language AI with Hybrid Architecture. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/falcon-h1-arabic (accessed 14 October 2026)
[12] Anwar, M., Freihat, A. A., Ibrahim, G., et al. (2026). Jais 2: A Family of Arabic-Centric Open Large Language Models. arXiv:2608.13580. https://arxiv.org/abs/2608.13580
[13] Fanar Team, Abbas, U., Ahmad, M. S., et al. (2026). Fanar 2.0: Arabic Generative AI Stack. arXiv:2603.16397. https://arxiv.org/abs/2603.16397
[14] Alkaabi, O., Alzubaidi, A., Alobeidli, H., et al. (2026). Alyah ⭐️: Toward Robust Evaluation of Emirati Dialect Capabilities in Arabic LLMs. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/emirati-benchmarks (accessed 14 October 2026)
[15] Muneer, A., Saraf, R., Gundluru, R., et al. (2026). Introducing Falcon ASR. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/falcon-asr (accessed 14 October 2026)
[16] Al-Khalifa, S., Durrani, N., Al-Khalifa, H., et al. (2025). The Landscape of Arabic Large Language Models (ALLMs): A New Era for Arabic Language Technology. arXiv:2506.01340, accepted at Communications of the ACM. https://arxiv.org/abs/2506.01340
[17] Alzubaidi, A., Alsuwaidi, S., Boussaha, B. E. A., et al. (2025). Evaluating Arabic Large Language Models: A Survey of Benchmarks, Methods, and Gaps. arXiv:2510.13430. https://arxiv.org/abs/2510.13430
[18] Talafha, B., Kadaoui, K., Magdy, S. M., et al. (2024). Casablanca: Data and Models for Multidialectal Arabic Speech Recognition. Proceedings of EMNLP 2024. https://aclanthology.org/2024.emnlp-main.1211/
[19] Alharbi, S., Alowisheq, A., Tüske, Z., et al. (2024). SADA: Saudi Audio Dataset for Arabic. Proceedings of ICASSP 2024, pp. 10286-10290. https://doi.org/10.1109/ICASSP48485.2024.10446243
[20] Sengupta, N., Sahu, S. K., Jia, B., et al. (2023). Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models. arXiv:2308.16149. https://arxiv.org/abs/2308.16149
[21] Huang, H., Yu, F., Zhu, J., et al. (2024). AceGPT, Localizing Large Language Models in Arabic. Proceedings of NAACL 2024. https://aclanthology.org/2024.naacl-long.450/
[22] Fanar Team, Abbas, U., Ahmad, M. S., et al. (2025). Fanar: An Arabic-Centric Multimodal Generative AI Platform. arXiv:2501.13944. https://arxiv.org/abs/2501.13944
[23] Cohere Labs (2025). c4ai-command-r7b-arabic-02-2025 model card. Hugging Face. https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025 (accessed 14 October 2026)
[24] Boussaha, B., Alyafeai, M., Alzubaidi, A., et al. (2025). Falcon-Arabic: A Breakthrough in Arabic Language Models. Hugging Face blog, Technology Innovation Institute. https://huggingface.co/blog/tiiuae/falcon-arabic (accessed 14 October 2026)
[25] Bari, M. S., Alnumay, Y., Alzahrani, N. A., et al. (2025). ALLaM: Large Language Models for Arabic and English. Proceedings of ICLR 2025 (arXiv:2407.15390). https://arxiv.org/abs/2407.15390
[26] Cohere Labs (2026). Aya. Cohere Labs research page. https://cohere.com/research/aya (accessed 14 October 2026)
[27] OpenAI (2025). Multilingual MMLU Benchmark Results. openai/simple-evals, GitHub. https://github.com/openai/simple-evals/blob/main/multilingual_mmlu_benchmark_results.md (accessed 14 October 2026)
[28] Anthropic (2026). Multilingual support. Claude documentation. https://platform.claude.com/docs/en/build-with-claude/multilingual-support (accessed 14 October 2026)
[29] Koto, F., Li, H., Shatnawi, S., et al. (2024). ArabicMMLU: Assessing Massive Multitask Language Understanding in Arabic. Findings of ACL 2024. https://aclanthology.org/2024.findings-acl.334/
[30] Al-Matham, R., Darwish, K., Al-Rasheed, R., et al. (2025). BALSAM: A Platform for Benchmarking Arabic Large Language Models. Proceedings of ArabicNLP 2025 (arXiv:2507.22603). https://aclanthology.org/2025.arabicnlp-main.21/
[31] Sajjad, H., Abdelali, A., Durrani, N., et al. (2020). AraBench: Benchmarking Dialectal Arabic-English Machine Translation. Proceedings of COLING 2020, pp. 5094-5107. https://aclanthology.org/2020.coling-main.447/
[32] Mousi, B., Durrani, N., Ahmad, F., et al. (2025). AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs. Proceedings of COLING 2025. https://aclanthology.org/2025.coling-main.283/
[33] Chen, Y., Liu, J., Tu, Y., et al. (2026). Habibi: Laying the Open-Source Foundation of Unified-Dialectal Arabic Speech Synthesis. arXiv:2601.13802. https://arxiv.org/abs/2601.13802
[34] Al Ali, M. K. and Aldarmaki, H. (2024). Mixat: A Data Set of Bilingual Emirati-English Speech. Proceedings of SIGUL 2024 at LREC-COLING 2024, pp. 222-226. https://aclanthology.org/2024.sigul-1.26/
[35] Radford, A., Kim, J. W., Xu, T., et al. (2022). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356. https://arxiv.org/abs/2212.04356

## Cite this paper

```bibtex
@techreport{qantara2026arabiclanguagemodels,
  title       = {Arabic language models for Gulf operations: a 2026 survey},
  author      = {Qantara Research},
  institution = {Qantara},
  year        = {2026},
  month       = {oct},
  url         = {https://qantara.systems/insights/arabic-language-models}
}
```

## Disclosure

Every model description, score, size, licence and date is quoted from the cited sources as read on 14 October 2026; dates without a stated release are Hugging Face repository creation dates, and "no public weights found" records our own search that day. The selection of models and benchmarks, Figure 2, the sum of ArabicMMLU's Gulf questions and the ranges drawn from published tables are ours; Table 3 and the workflow examples are illustrative. We ran no models for this survey.

Canonical: https://qantara.systems/insights/arabic-language-models