Best Arabic AI Models in 2026: ALLaM, Falcon and Beyond

How to Read Arabic AI Benchmarks Before You Compare Models

Almost every Arabic model announcement leans on the Open Arabic LLM Leaderboard (OALL), hosted on Hugging Face. Its first version attracted more than 700 model submissions from over 180 organisations. In February 2025 the maintainers released OALL v2, which removed saturated and machine-translated tasks because of their lower quality and possible cultural bias, and moved to native or human-curated tests: native Arabic MMLU, a human-translated MMLU, AlGhafa, EXAMS, MadinahQA, AraTrust, Belebele and ALRAGE, a new task that tests retrieval-augmented generation. When a vendor says its model tops the Arabic leaderboard, this is usually the leaderboard they mean.

ArabicMMLU, published at ACL 2024, is the knowledge test inside many of those scores. It contains 14,575 multiple-choice questions across 40 tasks, drawn from school exams in eight countries including Saudi Arabia, the UAE and Kuwait, and its authors report that more than half the questions are tailored to Arabic-specific contexts. When it was released, the best Arabic-centric model scored only 62.3%. It is a strong test of Modern Standard Arabic knowledge; it does not measure how well a model understands a customer writing in Najdi or Hijazi.

Saudi institutions have built their own measurement. BALSAM, launched by the King Salman Global Academy for Arabic Language and SDAIA at the Global AI Summit in September 2024, is a community benchmark whose paper describes 78 tasks across 14 categories and around 52,000 examples, with blind test sets to reduce the risk that models were trained on the answers. For dialect, academic work is catching up: the Absher benchmark (2025) uses more than 18,000 questions on Saudi dialects and found notable gaps in cultural inference and contextual understanding, even among strong models.

The practical lesson is simple. Leaderboard numbers are a shortlist tool, not a buying decision. Most published scores are self-reported by the model's developer, use multiple-choice formats in formal Arabic, and say little about conversation quality, dialect handling, or how a model behaves when grounded in your own documents. Use benchmarks to cut ten candidates to three, then test those three on your own data.

ALLaM and HUMAIN: Saudi Arabia's Own Arabic Model

ALLaM began at SDAIA's National Center for Artificial Intelligence. SDAIA built 7B, 13B and 70B parameter versions; ALLaM 13B became available to IBM watsonx users in May 2024, and in September 2024 Microsoft announced general availability of ALLaM 7B on Azure at the Global AI Summit in Riyadh. In February 2025 an instruction-tuned ALLaM-7B-Instruct-preview was published on Hugging Face under the Apache 2.0 licence, giving developers a Saudi-built Arabic model they could download and run themselves.

The model card for that open 7B model is unusually detailed. It was trained from scratch on about 4 trillion English tokens followed by 1.2 trillion mixed Arabic and English tokens, tuned on roughly 7 million instructions, and reports 67.78 on Arabic MMLU (zero-shot). Its limits are equally clear: a 4,096-token context window, which is short for long documents, and a size that trails the newest flagship models on reasoning. It is best suited to focused tasks, fine-tuning experiments, and on-premises deployments where weights must stay in-house.

In May 2025 HUMAIN, the PIF-backed national AI company, took ownership of the ALLaM brand and project in collaboration with SDAIA. In late August 2025 it launched HUMAIN Chat on web, iOS and Android, powered by ALLaM 34B, which HUMAIN describes as built and trained from scratch as an Arabic-centric foundation model. The app offers real-time web search, Arabic speech input across multiple dialects, switching between Arabic and English in one conversation, and hosting on HUMAIN infrastructure inside the Kingdom in line with the Personal Data Protection Law. ALLaM 34B has not been released as open weights.

Access for businesses is widening. On 26 August 2026 Microsoft and HUMAIN announced a long-term collaboration in which they intend to make ALLaM models available through Microsoft Foundry and to let organisations build Arabic-language agents in Microsoft 365 Copilot. As of this writing these were stated plans rather than general availability, so a buyer should confirm current access routes, hosting regions and terms directly before designing around them.

Falcon Arabic and Falcon-H1 Arabic: The Leaderboard Leaders

The Technology Innovation Institute (TII) in Abu Dhabi entered Arabic directly in May 2025 with Falcon Arabic, a 7B model built on Falcon 3-7B and trained on native Arabic data rather than machine translation, covering Modern Standard Arabic and regional dialects. TII reported that it outperformed Arabic models in its size range on OALL v2, including models several times larger, and it supported a 32,000-token context. At launch, TII offered it through its Falcon chat platform rather than as a standard open download.

On 5 January 2026 TII followed with Falcon-H1 Arabic, a family of 3B, 7B and 34B models. The architecture is a hybrid: each block runs Mamba state-space layers and Transformer attention in parallel, a design that handles long inputs efficiently. The 7B and 34B models support 256,000-token contexts and the 3B model 128,000, which matters for anyone feeding long contracts, policies or manuals into a model. TII says the training data was expanded to cover Gulf, Egyptian, Levantine and Maghrebi dialects.

On scores, TII reports Open Arabic LLM Leaderboard averages of 61.87% for the 3B model, 71.47% for the 7B and 75.36% for the 34B, and says the 34B model outperforms much larger systems such as Llama-3.3-70B. On TII's own dialect evaluation (AraDice, covering Egyptian, Gulf and Levantine), the scores are noticeably lower, in the 50s, than on the formal-Arabic tasks. That gap is honest and useful: it confirms that dialect remains harder than MSA, even for the current leaderboard leader.

For buyers, Falcon-H1 Arabic is the strongest option where long documents and benchmark accuracy per parameter matter. The key due-diligence item is licensing: TII's launch coverage pointed users to its public playground, and the terms attached to downloadable weights should be read carefully on TII's own model pages before you plan to self-host or fine-tune.

Jais 2, Fanar and the Wider Regional Field

Jais has been the region's best-known open Arabic model family since Inception, a G42 company, open-sourced the original 13B model with MBZUAI and Cerebras in 2023. In December 2025 Inception, Cerebras and MBZUAI's Institute of Foundation Models released Jais 2, a 70-billion-parameter model with a redesigned architecture and what its developers describe as the largest Arabic-first dataset they have assembled. Jais-2-8B-Chat and Jais-2-70B-Chat are available on Hugging Face, alongside a Jais Chat web app and mobile apps.

The Jais 2 70B model card lists the Apache 2.0 licence (access requires accepting conditions on Hugging Face), an 8,192-token context, support for Modern Standard Arabic, regional dialects and Arabic-English code-switching, and an AraGen score of 70.71. That combination of large size, permissive licence and explicit attention to code-switching makes Jais 2 one of the strongest candidates for organisations that want a fully self-hosted Arabic model, provided they have the GPU capacity a 70B model requires.

Qatar's Fanar, from the Qatar Computing Research Institute at Hamad Bin Khalifa University, is another credible option. Fanar-1-9B-Instruct was continually pretrained from Google's Gemma-2-9B on about 1 trillion Arabic, English and code tokens, supports MSA plus Gulf, Levantine and Egyptian Arabic, and is released under Apache 2.0 with a 4,096-token context. It is a sensible mid-size model for teams that want an open Arabic model that runs on modest hardware.

International vendors have also built regional models. Mistral AI released Mistral Saba in February 2025, a 24B model trained on curated data from the Middle East and South Asia, available through Mistral's API or deployable on a customer's own infrastructure on a single GPU. It is a useful reference point for buyers who want a compact regional model from a commercial vendor with enterprise support.

How Global Models Like GPT, Claude, Gemini and Qwen Handle Arabic

The global frontier models are strong in Modern Standard Arabic, and their developers publish numbers to prove it. On a version of MMLU professionally translated into Arabic, OpenAI's published results show o3-high scoring 0.904 and GPT-4.1 scoring 0.844. Anthropic's documentation reports Claude Sonnet 4.5 reaching 97.2% of its English performance in Arabic on the same kind of test. Google's Gemini models also handle Arabic well in practice, and all three vendors offer large context windows, tool use and mature developer platforms.

Among open-weight global models, Alibaba's Qwen3 family, released in April 2025 under Apache 2.0 for its dense models, lists 119 languages and dialects, and explicitly names Arabic varieties including Standard, Najdi, Levantine, Egyptian and Moroccan. That makes Qwen-class models a frequent base for Arabic fine-tuning, and a common comparison point in Arabic model papers: TII, for example, benchmarks Falcon-H1 Arabic against Qwen3 and Llama models of similar size.

There are two important caveats. First, translated MMLU measures knowledge in formal Arabic; it says nothing about a customer writing in Saudi dialect, mixing English product names into Arabic sentences, or expecting culturally appropriate phrasing. Second, the closed frontier models are usually consumed as APIs hosted outside the Kingdom, which brings cross-border data transfer obligations under Saudi Arabia's Personal Data Protection Law into play whenever personal data is involved.

In practice, many Saudi deployments combine both worlds: a global model for complex reasoning on non-sensitive content, and an Arabic-focused or self-hosted model for customer-facing conversations and personal data. The right mix depends on the workload, not on loyalty to one vendor.

Dialects, Open Weights and Data Residency

Dialect is where Arabic models still differ most. Most benchmarks test Modern Standard Arabic, but Saudi customers write in Najdi, Hijazi, Gulf and mixed registers, often with English terms in between. TII's own dialect scores for Falcon-H1 Arabic sit well below its formal-Arabic scores, and the Absher study found notable gaps on Saudi dialect and culture across the models it tested. A 2025 study that fine-tuned ALLaM-7B on Hijazi and Najdi data showed how much targeted work helps: the share of outputs classified as Saudi dialect rose from about 48% to 84%. Dialect performance can be improved, but it has to be measured on your own data.

Licensing determines what you can actually do with a model. ALLaM-7B-Instruct-preview, Jais 2, Fanar-1-9B and Qwen3's dense models are published under Apache 2.0, which permits commercial use, modification and self-hosting. ALLaM 34B is accessed through HUMAIN's products and partners rather than as downloadable weights. Falcon models carry TII's own licence terms. Closed frontier models such as GPT, Claude and Gemini are available only through their vendors' APIs and cloud partners.

Data residency is the third filter, and in Saudi Arabia it is not optional. The Personal Data Protection Law applies to any assistant that handles customer names, phone numbers, account details or health information. SDAIA's Regulation on Personal Data Transfer Outside the Kingdom, updated in September 2024, requires appropriate safeguards such as standard contractual clauses or binding common rules where a destination lacks an adequacy decision, and a risk assessment in certain cases. Hosting the model and the data inside the Kingdom removes a whole layer of that analysis.

Put together, these three factors often narrow the choice faster than any benchmark. A government-facing service handling personal data in Saudi dialect will usually land on an Arabic-focused model hosted in the Kingdom. An internal research tool working on public English and Arabic documents has far more freedom.

How a Saudi Business Should Choose an Arabic AI Model

Start with the job, not the model. Write down the three or four tasks the assistant must perform, such as answering policy questions, booking appointments, summarising Arabic documents or handling complaints, and collect 100 to 200 real examples of each from your own channels, including dialect, spelling mistakes and mixed Arabic-English text. This becomes your private test set, and it will tell you more than any public leaderboard.

Next, apply the hard filters: data residency, licence and budget. If personal data is involved and must stay in the Kingdom, your shortlist is self-hosted open models (ALLaM-7B, Jais 2, Fanar, Qwen-class models) or providers hosting inside Saudi Arabia. Cost follows from size: a 7B to 9B model can run on a single modern GPU, a 34B model needs considerably more memory, and a 70B model typically needs several GPUs or aggressive quantisation. API models have no hardware cost but bill per token, which grows quickly with long Arabic documents.

Then remember that most business accuracy comes from grounding, not from the base model. Retrieval-augmented generation (RAG) feeds the model the relevant passages from your own policies, product data and FAQs at answer time, so the assistant cites your documents rather than guessing. OALL v2 added ALRAGE precisely because this skill matters. A mid-size Arabic model with good retrieval, clean Arabic text extraction and well-structured documents will often beat a much larger model answering from memory.

Finally, run a blind comparison of two or three finalists on your test set, scored by native Arabic speakers from your target audience on correctness, dialect understanding, tone and refusal behaviour. Keep the architecture model-agnostic so you can switch when a better model appears; in this market, that happens every few months. Measure after launch too: track unanswered questions, escalations and customer satisfaction in Arabic separately from English.

Frequently Asked Questions

Conclusion

The Arabic AI market in September 2026 offers real choice. ALLaM gives Saudi Arabia a home-grown model family, from an open 7B release to HUMAIN's 34B flagship hosted in the Kingdom. Falcon-H1 Arabic leads the reported Open Arabic LLM Leaderboard numbers with very long context. Jais 2 and Fanar offer permissively licensed open weights, and the global frontier models remain strong in formal Arabic.

For a business, the decision comes down to a short list of practical questions: does the model understand your customers' dialect, can you legally host it where your data must live, can you afford it at real volume, and does it answer accurately when grounded in your own documents? Answer those on your own data, keep your architecture flexible, and the right Arabic model will be the one that serves your customers well, whichever leaderboard it tops.