Arabic LLMs in 2026: which you can download, and the GPUs to run them

Part of our AI models guide: 11 posts on choosing and running AI models.
Not every Arabic model you read about can be downloaded. TII's Arabic Falcon models (Falcon-H1-Arabic, Falcon-Emirati) are chat-only; only its general Falcons have public weights. The Gulf's Arabic models you can download and run are Jais 3 (UAE), ALLaM 7B (Saudi Arabia) and Fanar 2 (Qatar), all under Apache 2.0, a licence that allows commercial use. None of them was listed on OpenRouter, a marketplace for hosted AI models, when we checked on 19 November 2026, so teams usually run them on their own or rented GPUs: one GPU for a 7B model, two H100s for Jais-3-32B and one 8-GPU H200 server for the 375B Jais 3. On Jais 3's own published tests, global open models such as Qwen3.5 still score higher on Arabic chat, while Jais 3 leads on dialects and translation, so test both on your own prompts.
Open weights means you can download a model's weights, the numbers it learned in training, usually from Hugging Face, the main site for sharing AI models, and run it yourself (see open-weight vs closed models). Size is in billions of parameters (B); context is how much text a model reads at once, in tokens, the smallest units of text an AI model reads and writes, often a short word or part of one, and the unit AI services bill by.
Gulf Arabic AI models in one table
Model | Maker, country | Released | Sizes and context | Licence | Download or chat only? |
|---|---|---|---|---|---|
Jais 3 | MBZUAI and Inception, UAE | 15 Oct 2026 | Six sizes, 0.9B to 375B (MoE, 23B active); 512K context; Arabic and English | Apache 2.0 | Download, plus training checkpoints and a 23.8B-token data sample |
Jais 2 | Inception and MBZUAI, UAE | 9 Dec 2025 | 8B and 70B; 8,192 context; Arabic (standard and dialects) and English | Apache 2.0 | Download, after accepting conditions |
K2 Horizon | MBZUAI, UAE | 3 Sep 2026 | Same six sizes as Jais 3; 512K context; cards list English | Apache 2.0 | Download, with training data and code |
Falcon-H1-Arabic, Falcon-Emirati | TII, UAE | 5 Jan and 6 Oct 2026 | 3B, 7B, 34B, up to 256K context; Emirati 7B | None published | Chat only (Falcon Chat and demos) |
Falcon3, Falcon-H1, Falcon-H1R | TII, UAE | Dec 2024 to Jan 2026 | General models, 0.5B to 34B; Falcon-H1R is a 7B reasoning model | TII Falcon licence | Download |
Falcon 7B and 40B | TII, UAE | 2023 | 7B and 40B | Apache 2.0 | Download |
ALLaM-7B-Instruct-preview | SDAIA, now published by HUMAIN, Saudi Arabia | Feb 2025 (preview) | 7B; 4,096 context; Arabic and English | Apache 2.0 | Download |
ALLaM 34B | HUMAIN, Saudi Arabia | Aug 2025 | 34B | None published | Chat only (HUMAIN Chat) |
humain-m3 | HUMAIN with China's MiniMax, Saudi Arabia | Preview, 3 Sep 2026 | 428B MoE, trained further on more than 1T Arabic tokens | MiniMax Community Licence (planned) | Not yet: no public weights on Hugging Face |
Fanar 2 (Fanar-2-27B-Instruct) | QCRI, Qatar | Announced 9 Dec 2025; weights Mar 2026 | 27B, built on Google's Gemma 3; 32K context | Apache 2.0 (check Gemma terms too) | Download |
The Gulf builds its own models because Arabic and its dialects are thin in global training data, and governments want AI they control. In the UAE, the Technology Innovation Institute (TII) makes Falcon, and MBZUAI, an Abu Dhabi AI university, makes K2, which it calls fully open. Inception, part of Abu Dhabi's G42 tech group, builds Jais with MBZUAI; Jais 3 is built on K2 Horizon. Saudi Arabia's ALLaM comes from SDAIA, the Saudi Data and AI Authority, and is now published by HUMAIN, the national AI company. Qatar's Fanar comes from QCRI, the Qatar Computing Research Institute.
The 375B models are mixture-of-experts (MoE): only about 23B parameters work on each token, so they answer quickly for their size, but all 375B must sit in GPU memory.
The 2023 and 2024 Jais models (up to 70B) are still on Hugging Face; Jais 2 and 3 succeed them. QCRI has said Fanar 3 is due in December 2026. For US and Chinese open models and the GPUs they need, see how many GPUs it takes to run an LLM.
Falcon: download or chat only
This is the main trap. Some comparison sites suggest running "Falcon Arabic" yourself, but TII has not published its Arabic models: on 19 November 2026 its Hugging Face page listed 137 models, none Arabic or Emirati. Falcon Arabic, Falcon-H1-Arabic and Falcon-Emirati run only on Falcon Chat and demos. TII reports 75.36% for Falcon-H1-Arabic-34B on the Open Arabic LLM Leaderboard (vendor-reported), but you cannot host it. Falcon-H1R, which you can download, lists English and multilingual, so it is not an Arabic specialist.
GPU requirements for each model
Using these models in a product usually means running them on bought or rented GPUs; see our rent, buy or API guide.
In BF16, the standard 16-bit format, each parameter takes about 2 bytes; FP8, an 8-bit format, roughly halves that. Our rule of thumb: keep the weights to about 80% of GPU memory, leaving the rest for the KV cache (the model's working memory per conversation); long contexts and many users need more. See our GPU memory guide.
Model | Weights in BF16 | Minimum GPUs (sizing only) | Source | Watch for |
|---|---|---|---|---|
K2-Horizon-7B and Jais-3-7B (about 9B), ALLaM-7B, Falcon-H1R 7B | About 14 to 18 GB | One GPU, for example one H100 | Hugging Face sizes; K2 Horizon card for its 7B | Prompts near K2's and Jais's 512K-token limit, or long reasoning answers, need far more KV cache memory than the weights |
Fanar-2-27B-Instruct | About 54 GB | One H100 (80 GB) or one H200 | FlexiCloud estimate | 32K context |
Jais-3-32B | About 70 GB (34.8B parameters) | Two H100 (80 GB) | Jais-3-32B model card (BF16, split across two GPUs) | One H200, B200 or B300 may hold it; test first (our estimate) |
Jais-2-70B | About 144 GB | Two H200 or four H100; one H200 or B200 in FP8 | FlexiCloud estimate | The card's serving example splits it across two GPUs |
K2-Horizon-375B-A23B | About 760 GB (379B parameters, per Hugging Face) | One 8-GPU H200 server | K2 Horizon model card (BF16, split across eight GPUs) | All parameters must sit in memory, not just the 23B active ones |
Jais-3-375B-A23B | About 760 GB (our estimate) | One 8-GPU H200 server, or two 8-GPU H100 servers (16 GPUs) | Jais 3 model card | The FP8 version roughly halves the weights (our estimate); two servers need a fast network |
Arabic training on top of a large global model | Depends on the base model | A multi-server B300 cluster or a GB300 NVL72 rack (72 GPUs) | FlexiCloud guidance | How Fanar and humain-m3 were built |
We arrange whole 8-GPU servers (often called nodes), not single GPUs. One server can run several copies of a 7B to 32B Arabic model, or one 375B flagship in BF16 on H200, B200 or B300; B200 and B300 leave more memory for long Arabic documents. For two servers, see NVLink and InfiniBand explained and renting a multi-node B300 cluster.
Racks matter for continued pretraining (training a global model further on Arabic text) and for serving a 375B model to many users at long context. A GB300 NVL72 rack joins 72 GPUs as one; see what NVL72 means and our GB300 NVL72 rack terms. For smaller fine-tunes (extra training on your own data), see what fine-tuning on B300s costs.
Licences: what each model lets you do
Most Gulf models use Apache 2.0, which allows paid products and APIs. The TII Falcon licence is royalty-free but has conditions: TII can update its Acceptable Use Policy, and you must carry its limits into your own customer terms.
Licence | Models | Commercial use | Watch for |
|---|---|---|---|
Apache 2.0 | Jais 3, Jais 2, K2 Horizon, ALLaM-7B, Fanar 2, Falcon 7B and 40B | Allowed | Keep the licence and notices; for Fanar 2, read Google's Gemma terms too |
TII Falcon licence (Dec 2024) | Falcon3, Falcon-H1, Falcon-H1R | Allowed and royalty-free, with conditions | Derivatives must say they are "built using artificial intelligence technology from the Technology Innovation Institute" |
No public licence | Falcon Arabic, Falcon-H1-Arabic, Falcon-Emirati, ALLaM 34B | Cannot be self-hosted | Use the maker's chat service, or pick an Apache 2.0 model |
humain-m3, if released, is planned under the MiniMax Community Licence, where commercial use above US$20M yearly revenue needs MiniMax's written approval; read its own licence when it ships. For selling tokens under each licence, see the licence table in our GPU sizing guide, linked above. This is not legal advice; read each model's LICENSE file.
Do Gulf models beat global models in Arabic?
Not across the board. The table uses only the scores on Jais 3's model cards (October 2026), so every chat score uses the same benchmark version, AraGen v3. All figures are vendor-reported; an AI judge scores the translation test.
Test | Jais 3 | Global open model | Higher |
|---|---|---|---|
AraGen v3, Arabic chat quality | Jais-3-375B: 72.4 | Qwen3.5-397B: 78.9 | Qwen3.5 |
AraGen v3, Arabic chat quality | Jais-3-32B: 71.0 | Gemma-4-31B: 73.4 | Gemma 4 |
QADI, Arabic dialect identification | Jais-3-375B: 71.1 | Qwen3.5-397B: 62.3 | Jais 3 |
Islamic questions and answers | Jais-3-375B: 90.2 | Qwen3.5-397B: 88.0 | Jais 3 |
FLORES-200 translation, AI-judged | Jais-3-375B: 91.9 | Qwen3.5-397B: 83.6 | Jais 3 |
Falcon's and Jais 2's leaderboard scores come from different tests, so do not compare them with these. On these vendor-reported scores, Jais 3 leads on dialects, Islamic questions and translation, while Qwen3.5 and Gemma 4 lead on general Arabic chat: test one Gulf and one global model on your own Arabic prompts.
Falcon-7B (2023) is still the most downloaded Gulf model: about 126,000 Hugging Face downloads in the 30 days to mid-November. ALLaM-7B had about 30,700, K2-Horizon-7B about 24,500, and all Jais 3 repos together about 33,000. The counts leave out many third-party copies.
Why Arabic can cost more tokens
Models read, write and bill by the token (see AI token pricing explained). Tokenizers, which cut text into tokens, were mostly built on English, so they often split Arabic into more pieces: an Arabic answer costs more on per-token APIs and takes more GPU time when self-hosted.
Tokenizer | Arabic tokens for the same content, against English | Study |
|---|---|---|
GPT-4 | About 3 times | Petrov et al., 2023, FLORES-200 sentences (ratio our arithmetic) |
GPT-4o | 1.44 times | August 2026 preprint: an early study of 120 Python-tutoring items, Arabic machine-translated |
Qwen2.5-7B | 1.70 times | Same preprint |
Mistral-7B | 3.86 times | Same preprint |
Arabic-first tokenizers, such as Jais's, narrow the gap. On your own GPUs, fewer tokens per answer means more answers per GPU-hour: a capacity gain, not just a smaller bill.
Data rules and where to run
Gulf data rules are mostly set by sector, not one blanket law. Saudi Arabia's personal data law does not yet require personal data to stay in the Kingdom. Draft amendments, open for comment from 6 October to 5 November 2026, would require local storage while still allowing transfers abroad; they are not final. Sector regulators, for example in banking, can be stricter. The UAE's personal data law allows transfers abroad under conditions, and sector rules, for example for health data, can require local storage.
So regulated work for government, banks and health usually needs GPUs in the right country; other work can run in a nearby region. Your own GPUs keep prompts and answers away from a model maker's API; see bare metal vs cloud GPUs. Not legal advice: check with your regulator or lawyer.
The Gulf AI build-out and US export rules
Sovereign funds pay for the build-out, which depends on US approval to import NVIDIA's top chips. These are national projects and policy, not places to rent GPUs.
In the table, A:5 is the US group of close allies with the lightest export controls, and D:5 is its arms-embargo group, which includes China. TPP (total processing performance) is the US measure of chip power: 12 million TPP is about 760 H100 GPUs at roughly 15,800 TPP each (our arithmetic).
Date | Where | What happened | Status on 19 Nov 2026 |
|---|---|---|---|
22 May 2025 | UAE | Stargate UAE: a 1 GW cluster of NVIDIA GB300 systems in a planned 5 GW Abu Dhabi campus, built by G42 with OpenAI and US partners | First 200 MW targeted for 2026; no public confirmation it is running |
From May 2025 | Saudi Arabia | HUMAIN, owned by the Public Investment Fund, forms chip partnerships with NVIDIA, AMD and Qualcomm | Targets about 3 GW of data-centre capacity by 2030; a later NVIDIA phase is planned on Vera Rubin |
19 Nov 2025 | UAE and Saudi Arabia | US Commerce approves G42 and HUMAIN to buy the equivalent of up to 35,000 GB300 chips (widely reported as each company's limit), with security and reporting conditions | Approved |
9 Dec 2025 | Qatar | Qai, a Qatar Investment Authority company, and Brookfield form a US$20bn AI infrastructure joint venture | No site or start date announced |
10 Jul 2026 | UAE | The US Bureau of Industry and Security (BIS) moves the UAE to Country Group A:5; G42 group companies may get advanced chips without a licence | In force; G42's approval ends 6 Apr 2027 unless BIS gives further notice |
16 Nov 2026 | Worldwide | BIS remote-access rule: a licence is needed to supply advanced chips to any data centre giving more than 12 million TPP to companies headquartered in China, other D:5 countries or Macau (ultimate parents count too) | In force from 19 Nov 2026, when it was published in the Federal Register; 90 days for existing arrangements |
Dedicated servers for Jais and Fanar
Because Jais 3 and Fanar 2 were not on OpenRouter when we checked, putting them into production usually means running them yourself, on your own or rented GPUs. Our six regions include the Gulf, where we source dedicated 8-GPU H100, H200, B200 and B300 servers through verified operators, plus GB300 NVL72 racks on request for continued pretraining. Terms start at one year, billed upfront for the whole term; we neither rent by the hour nor operate the machines. Before signing anything, run through our questions to ask a GPU provider.
Tell us the model, context length and traffic, and we will price a Gulf server for it.
Where these figures come from
Checked 19 November 2026 unless dated.
- Hugging Face model cards: Jais 3, Jais-3-32B, Jais 2, K2 Horizon, Falcon-H1R, ALLaM-7B and Fanar 2; model lists, parameter counts and 30-day downloads from Hugging Face's API.
- TII: Falcon-H1-Arabic and launch release (5 January 2026), Falcon-Emirati (6 October 2026), TII Falcon licence (December 2024).
- Middle East AI News: Jais 3 (15 October 2026), ALLaM 34B (25 August 2025), humain-m3 (3 September 2026), Fanar 3 plan (December 2025). The Peninsula on Fanar 2 (10 December 2025). MiniMax M3 licence. OpenRouter model list.
- Tokens: arXiv 2608.09046 (August 2026); Petrov et al. (NeurIPS 2023).
- Data rules: Clyde & Co on Saudi PDPL drafts (October 2026); UAE government portal on data protection laws (updated 4 December 2025).
- Stargate UAE: Gulf News (22 May 2025), The National (5 December 2025), Khaleej Times (13 August 2026). HUMAIN: NVIDIA (13 May 2025), The National (27 October 2025), AMD (31 August 2026) and The National (22 October 2026). Qai: QNA (9 December 2025).
- Export rules: The National (20 November 2025), Federal Register 2026-14132 (14 July 2026), BIS press release (16 November 2026) and the rule text, Federal Register document 2026-23529 (published 19 November 2026).