Arabic-First LLMs: When Regional Models Beat Global Ones

The question is narrower than "which model is best at Arabic"

The Gulf now has a credible line-up of Arabic-first language models. The UAE has Jais 2 from Inception, MBZUAI and Cerebras, and Falcon-H1 Arabic from the Technology Innovation Institute (TII). Saudi Arabia has ALLaM, first built at SDAIA and now carried forward by HUMAIN, which in September 2026 also previewed humain-m3, a 428-billion-parameter model developed with MiniMax. Each launch came with a claim to lead Arabic AI. Several of those claims are true on the benchmark that was chosen to support them.

For an enterprise, "which model is best at Arabic" is the wrong question. Our argument is that a regional model beats a global one under four specific conditions: your users write in dialect or mix Arabic and English; you need the weights inside your own environment or the country; you process enough Arabic text that tokens per document drive cost; or the content is culturally sensitive enough that alignment matters more than raw reasoning. Outside those conditions, a strong global model with good Arabic retrieval is often the better and simpler choice for formal Modern Standard Arabic (MSA) work.

Public leaderboards will not settle this for you. They mostly measure multiple-choice questions in MSA. The evidence below shows that the model leading on MSA exams is often not the one leading on dialects, and that dialect scores are low for every model. The only reliable referee is a golden set built from your own Arabic traffic.

Model availability, licences and scores in this article are as of 27 September 2026. Most figures come from technical reports and announcements written by the model builders, and are labelled as such.

Why Arabic is hard for general-purpose models

Three features of Arabic separate it from the languages most LLMs were optimized for.

One written standard, many spoken languages. MSA is the language of government, news and contracts. Daily life runs in dialects (Gulf, Najdi, Egyptian, Levantine, Maghrebi and others) that differ in vocabulary, morphology and syntax, and are often written informally, sometimes in Latin script ("Arabizi"), and often mixed with English mid-sentence. A customer-service chat in Riyadh and a ministerial circular are different tasks. The BALSAM benchmark authors attribute Arabic's lag behind English to "data scarcity, linguistic diversity of Arabic and its dialects, morphological complexity".

Rich morphology. Arabic attaches prefixes, suffixes and clitics to roots, so one written word can carry what English expresses in several. Tokenizers trained mostly on English break these words into many fragments.

Culture is part of the answer. Questions about religion, heritage, local law, names and Hijri dates have answers that a model trained largely on English web text may get subtly wrong, even when its grammar is perfect.

The tokenizer tax

Tokenization is where the language challenge becomes a line item. Every API bill, latency figure and context window is counted in tokens, and Arabic needs more of them.

The original Jais report (2023) showed the extreme case: the English-centric GPT-2 tokenizer had a fertility score of 4.17 on Arabic text against 1.10 on English, while Jais's bilingual tokenizer scored 1.05 on Arabic. Modern commercial tokenizers are much better. An August 2026 audit measured Arabic at 1.44 times the English token count with GPT-4o's tokenizer, 1.70 times with Qwen2.5 and 3.86 times with Mistral-7B. That study used a small, machine-translated technical corpus and treats Arabic as exploratory, so read the numbers as a direction, not a constant.

Three practical consequences follow:

• A token premium is a price premium. At 1.44 times, the same Arabic content costs about 44% more than English at the same per-token price, and fills a context window 44% faster.
• Compare cost per document, not price per token. A model with an Arabic-efficient tokenizer can be cheaper per document even at a higher per-token price. The only way to know is to tokenize your own documents with each candidate.
• Open models vary most. If you plan to self-host, the tokenizer of the base model you choose matters as much as its parameter count.

The regional model landscape in September 2026

The regional models differ less in headline quality than in how much control they give you.

Jais 2 (UAE). Released in December 2025 by Inception (a G42 company), MBZUAI and Cerebras, with a technical report in July 2026. It comes in 8B and 70B sizes, trained from scratch on Arabic, English and code with a new 150,272-token vocabulary designed for MSA, dialects and code-switching. The model card lists an Apache 2.0 licence. The constraint is the context window: 8,192 tokens, which limits long-document retrieval and multi-step agent work.

Falcon-H1 Arabic (UAE). Launched by TII on 5 January 2026 in 3B, 7B and 34B sizes on a hybrid Mamba-Transformer architecture, with context windows up to 256,000 tokens. TII's blog describes about 300 billion tokens of further training on top of the Falcon-H1 base, with wider coverage of Egyptian, Levantine, Gulf and Maghrebi dialects. TII claims the top position on the Open Arabic LLM Leaderboard (OALL), reporting about 75% for the 34B model. At the time of writing we could not confirm the licence terms of the weights from TII's announcement, so check them before planning self-hosting.

ALLaM (Saudi Arabia). SDAIA's 2024 technical report described a family built mostly by continued training of existing models. The openly published ALLaM-7B-Instruct-preview, now under HUMAIN's account on Hugging Face, was trained from scratch on 4 trillion English tokens followed by 1.2 trillion mixed Arabic and English tokens, under Apache 2.0 with a 4,096-token context. ALLaM 34B, launched with the HUMAIN Chat app on 25 August 2025, is described as hosted on HUMAIN's infrastructure in Saudi Arabia. We found no public weights or technical report for the 34B model.

humain-m3 (Saudi Arabia). Announced on 3 September 2026 and developed by MiniMax, a Chinese AI developer. HUMAIN describes a 428-billion-parameter mixture-of-experts model, further pretrained on more than one trillion tokens of Arabic-native content, in research preview on HUMAIN Node. HUMAIN says it achieved the highest average across seven public Arabic benchmarks among the frontier models it tested. The release does not name the benchmarks or the competitors, so treat this as an unverified vendor claim. HUMAIN plans to publish weights under the MiniMax Community License, targeted for October 2026.

The humain-m3 case carries a broader lesson. "Sovereign" and "Arabic-first" describe who controls training, alignment and hosting. They do not guarantee that the base architecture or weights originated in the region. Procurement teams should ask where the base model came from, under which licence, and which party can change it.

What the benchmarks show, and what they hide

Global models have been strong on standard Arabic

When ArabicMMLU was published in 2024, with 14,575 questions drawn from school exams across North Africa, the Levant and the Gulf, GPT-4 scored 72.5% against 62.3% for the best Arabic-centric model of the time, Jais-chat 30B. On exam-style MSA, a general frontier model led.

The regional models have since closed and in places reversed that gap among open models. In the Jais 2 report, Jais 2 70B averages 79.36% on the six OALL2 multiple-choice tasks, against 74.23% for Llama 3.3 70B and 71.20% for Qwen2.5 72B. These are the vendor's own runs.

Dialects change the ranking

The same report shows the ranking changing on dialects. On Dialectal ArabicMMLU, Qwen2.5 72B scores highest at 71.61%, ahead of TII's general Falcon-H1 34B at 69.54% and Jais 2 70B at 67.32%. The authors acknowledge the point themselves: dialect understanding remains hard even for large models, and the leaders differ from those on MSA benchmarks.

Generation in dialect is harder still. Falcon-H1 Arabic's own release reports AraDiCE dialect scores of roughly 50% across Egyptian, Gulf and Levantine for all three sizes, well below its scores on other benchmarks. In dialect translation, the Jais 2 report includes closed models (Gemini 2.5 Pro and Flash, GPT-5) and describes Jais 2 70B as the best open system that "in many cases approaches" the closed frontier models. That is a credible result for an open 70B model, but it is not a clean win.

INVISOM reading: regional models now lead open-model comparisons on MSA knowledge and culturally grounded tasks. Dialect performance is uneven for everyone, and frontier closed models remain competitive. A buyer who picks from an MSA leaderboard to serve dialect-speaking customers is choosing on the wrong evidence.

What public benchmarks do not measure

The second version of OALL, launched in February 2025, improved the field. It removed saturated and machine-translated tasks, added native ArabicMMLU, a safety set (AraTrust) and a retrieval task (ALRAGE), and found that a silent bug in one task had distorted earlier rankings. BALSAM, launched with SDAIA and the King Salman Global Academy for Arabic Language, adds 78 tasks with blind test sets to limit contamination. TII researchers went further in April 2026 with QIMMA, a leaderboard that validates benchmark quality before scoring any model.

The gaps that remain are the ones that matter most to enterprises. A 2025 survey of more than 40 Arabic benchmarks identified limited temporal evaluation, insufficient multi-turn dialogue assessment and cultural misalignment in translated datasets. None of the major suites tests your terminology, your forms, your mix of Arabic and English, or your cost per task. Some also depend on LLM judges: OALL2's ALRAGE is scored by Qwen2.5 72B, and research shows that judge models can recognize and favour their own outputs.

Sovereignty and hosting: where regional models are hardest to beat

For many Gulf organizations, the decision is made by control requirements before accuracy is measured. Three questions sort the options.

Must the data stay in-country? Data classification under the NDMO standards and personal-data rules in Saudi Arabia and the UAE often require that sensitive prompts and documents be processed within the country. A global model is only an option if the specific model version is served from an in-country region with no fallback to global processing. Confirm this per model with the provider in writing. Our guide to the NDMO standards covers how classification drives these decisions.

Must you hold the weights? Open-weight models (Jais 2 and ALLaM 7B under Apache 2.0, humain-m3 if its weights are released as announced) can run on infrastructure you control, be versioned, and be fine-tuned without depending on a vendor's roadmap. Our comparison of RAG, fine-tuning and custom models shows how quickly closed-model fine-tuning options can disappear.

Who can change the model under you? A hosted national model removes infrastructure work but adds a dependency on one provider's release schedule, terms and capacity. Treat it like any critical supplier: version pinning, deprecation notice, exit plan.

Self-hosting has a cost. A 70B model needs serious GPU capacity and an operations team. For moderate volumes, an in-country hosted endpoint, regional or global, may be cheaper than owning the hardware. The economics belong in the same evaluation as accuracy.

An evaluation protocol for Arabic: build the golden set first

The most valuable asset in this decision is not a model. It is a test set that reflects how your users actually write. This is how we recommend building one.

Illustrative sizing: a few hundred well-annotated items, with at least 50 per dialect or register that matters to you, usually separates candidates more reliably than thousands of generic benchmark questions. The point is coverage of your real mix, not volume.

Run every shortlisted model, regional and global, through the same harness with the same retrieval layer and prompts. Then compare results by register. A model that wins overall but fails on Gulf dialect may be the wrong choice for a customer channel and the right one for policy drafting. Many organizations will end up routing by task.

A decision guide

Two patterns appear most often in practice:

• Split by channel. A regional open-weight model, hosted in-country, handles citizen- or customer-facing dialect conversations and sensitive data. A frontier global model handles formal MSA analysis, long documents and English-heavy work where data rules allow.
• One model plus Arabic retrieval. Where usage is mostly formal MSA and data can use an in-country endpoint, a single strong global model with well-built Arabic retrieval (Arabic-aware chunking, normalization of letter variants and diacritics, and Arabic embeddings tested on your own content) is often enough.

Both patterns depend on separating knowledge from the model, as described in our guide to taking AI from experiments to enterprise value. If documents, retrieval and evaluation stay independent of any one model, switching from a global to a regional model, or back, becomes an evaluation run rather than a re-platforming project.

What to do in the next 90 days

1. Profile your Arabic. Sample a month of real requests and documents and measure the register mix: MSA, each dialect, Arabizi, code-switching. Most organizations do not know these numbers.
2. Build and freeze a golden set of a few hundred items following the protocol above, with native-speaker references and a hidden subset.
3. Shortlist four to six models: at least two Arabic-first (for example Jais 2 and Falcon-H1 Arabic, plus ALLaM or humain-m3 if available to you) and two global frontier models available in your hosting constraints.
4. Tokenize your own documents with each candidate and compute cost per document and per correct answer.
5. Write down the hosting requirement per use case with your data-governance and security teams: in-country processing, weight ownership, or neither.
6. Decide per channel, not per company, and record the evidence so the decision can be re-run when the next model is released.
7. Schedule a re-test for the first quarter of 2027. Several models in this article are in preview or recently released, and rankings will change.

Frequently asked questions

Are Arabic-first LLMs better than GPT or Gemini at Arabic?

Sometimes. In vendor-reported tests, Jais 2 70B leads open models on standard Arabic exams and culturally grounded tasks, while closed frontier models remain competitive on dialect translation. On dialect understanding, no model is consistently ahead. The answer depends on your dialect mix and tasks, so test on your own data.

Why does Arabic cost more to process with LLMs?

Tokenizers built mainly for English split Arabic words into more pieces. One 2026 audit measured Arabic at 1.44 times the English token count with GPT-4o's tokenizer and 3.86 times with Mistral-7B's. More tokens mean higher cost, more latency and less usable context.

Which benchmark should we trust for Arabic LLMs?

Use public benchmarks to build a shortlist, not to decide. The Open Arabic LLM Leaderboard, ArabicMMLU and BALSAM are useful for standard Arabic, but they cover dialects, multi-turn conversation and enterprise tasks only partly. Make the final decision on a golden set built from your own users' Arabic.

Do we need a regional model for data sovereignty in Saudi Arabia or the UAE?

Not always. If a global model version is processed entirely in-country under your data classification rules, it may qualify. If you must hold the weights or run fully on your own infrastructure, open-weight regional models such as Jais 2 or ALLaM 7B are the most direct route.

Sources and further reading

• Jais 2: A Family of Arabic-Centric Open Large Language Models, MBZUAI, Inception and Cerebras, July 2026.
• Jais-2-70B-Chat model card, Inception, Hugging Face, accessed 27 September 2026.
• Jais and Jais-chat technical report, Sengupta et al., 2023.
• Introducing Falcon-H1-Arabic, Technology Innovation Institute, 5 January 2026.
• Abu Dhabi's TII launches Falcon-H1 Arabic, Technology Innovation Institute, 5 January 2026.
• ALLaM-7B-Instruct-preview model card, HUMAIN, Hugging Face, accessed 27 September 2026.
• ALLaM: Large Language Models for Arabic and English, Bari et al., SDAIA, 2024.
• HUMAIN Chat application, Saudipedia, accessed 27 September 2026.
• HUMAIN unveils humain-m3, HUMAIN via PR Newswire, 3 September 2026.
• Saudi AI firm Humain unveils model based on China's MiniMax, Bloomberg, 3 September 2026.
• The Open Arabic LLM Leaderboard 2, Hugging Face, 10 February 2025.
• ArabicMMLU, Koto et al., Findings of ACL 2024.
• BALSAM: A Platform for Benchmarking Arabic Large Language Models, Al-Matham et al., 2025.
• AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs, Mousi et al., 2024.
• Evaluating Arabic Large Language Models: A Survey of Benchmarks, Methods, and Gaps, Alzubaidi et al., 2025.
• Are Arabic Benchmarks Reliable? QIMMA's Quality-First Approach, AlQadi et al., April 2026.
• Measuring the Tokenization Premium, Roy, Roy and Patel, August 2026.
• LLM Evaluators Recognize and Favor Their Own Generations, Panickssery, Bowman and Feng, 2024.