Coaley Peak GEO Research Centre Engine guide
Jais and ALLaM جيس وعلّام How the Gulf's Arabic-first AI models learn, and how a brand gets into them.
Jais, from the UAE, and ALLaM, from Saudi Arabia, are the Arabic-first language models built in the Gulf. Jais powers JAIS Chat and ALLaM powers HUMAIN Chat. Neither documents a crawler, so this guide concentrates on what they were trained on, how they answer, and how a brand makes its Arabic presence count.
Independent guide: Coaley Peak is not associated with Inception (G42), MBZUAI, SDAIA and HUMAIN. Trademarks
- Owner
- Inception (G42), MBZUAI, SDAIA and HUMAIN
- Launched
- Jais August 2023; ALLaM paper July 2024; HUMAIN Chat August 2025
- Home market
- The Gulf and the wider Arabic-speaking world
- Languages
- Arabic (Modern Standard and dialects) and English
- Open weights
- Every Jais generation and ALLaM 7B; not ALLaM 34B
- Crawlers
- None published; Common Crawl's CCBot fed Jais's training
Overview
What Jais and ALLaM is, and who uses it
Most AI assistants learn Arabic as one language among many. Jais and ALLaM were built for it. Jais was released in August 2023 by Inception, part of Abu Dhabi's G42 group, with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Cerebras Systems. ALLaM was developed by the National Center for Artificial Intelligence at the Saudi Data and AI Authority (SDAIA), and set out in a technical paper in July 2024.
Both now have consumer assistants. JAIS Chat, published by MBZUAI, runs on Jais 2 on the web at jaischat.ai and on iOS and Android, in Arabic and English. HUMAIN Chat comes from HUMAIN, the AI company launched by Saudi Arabia's Public Investment Fund in May 2025, which calls ALLaM its flagship model. HUMAIN Chat runs on ALLaM 34B, is hosted and operated in Saudi Arabia, and has a real-time web search that users can switch on.
Neither developer publishes usage figures, and neither documents a crawler, user agent or opt-out token. What is documented in unusual detail is training data: the Jais and ALLaM papers set out how much Arabic they used, where it came from and how it was filtered. For a brand, that makes these models a pre-training question first and a retrieval question second.
Arabic web content is scarce compared with English. The Jais team described its 72 billion Arabic tokens as the largest Arabic training corpus assembled at the time, and repeated it in training to make it go further. When the pool is that small, every substantial, well-written Arabic page about a brand counts for more.
Timeline
Jais and ALLaM, from launch to today
Saudi Arabia establishes SDAIA by royal decree, with the National Center for AI that later builds ALLaM.
Jais and Jais-chat, 13-billion-parameter Arabic-centric models, are released by Inception (G42), MBZUAI and Cerebras, trained on 395 billion tokens of Arabic, English and code.
Jais 30B is released, trained on 126 billion Arabic, 251 billion English and 50 billion code tokens.
MBZUAI publishes the JAIS Chat app on iOS.
SDAIA's technical paper on ALLaM, its Arabic and English language models, is published.
The Jais family is released: 20 models from 590 million to 70 billion parameters, trained on up to 1.6 trillion tokens, under Apache 2.0.
ALLaM-7B-Instruct-preview is published on Hugging Face, pre-trained on 5.2 trillion tokens.
HUMAIN launches under Saudi Arabia's Public Investment Fund, with ALLaM as its flagship model.
HUMAIN Chat, powered by ALLaM 34B, is released on the Saudi App Store.
HUMAIN unveils HUMAIN One, an Arabic-first operating system powered by ALLaM.
Jais 2, trained from scratch in 8-billion and 70-billion-parameter sizes with an Arabic-centric vocabulary, is published under Apache 2.0, and JAIS Chat moves to it.
HUMAIN Chat appears on Apple's UK App Store, after its Saudi launch.
Things to know
Things to know about Jais and ALLaM
Write for Modern Standard Arabic and the dialects. Jais 2 explicitly targets Modern Standard Arabic, regional dialects and Arabic–English code-switching, and HUMAIN Chat's voice input covers dialects from the Gulf to the Maghreb.
Neither assistant publishes a knowledge cut-off. Assume nothing recent is in the weights; HUMAIN Chat's optional web search is the only documented live route.
No crawler user agent is published by any of these developers, so server logs cannot attribute their visits and robots.txt cannot target them.
Jais 13B's training mix included the Arabic subset of C4 and the English Pile-CC, both filtered from Common Crawl. Being crawlable by CCBot was one route into the original Jais.
ALLaM's Arabic web data was crawled in-house and filtered hard. Documents below 95% confidence in language identification, under 30 words, duplicated, or high in spam or stop words were dropped, so short or mixed-language Arabic pages are unlikely to survive.
Open weights differ. Every Jais generation is Apache 2.0, though downloads ask for contact details; ALLaM 7B is tagged Apache 2.0; ALLaM 34B, which powers HUMAIN Chat, has no published weights.
HUMAIN Chat is hosted and operated in Saudi Arabia, and HUMAIN's executives cite compliance with the Saudi Personal Data Protection Law.
HUMAIN has said ALLaM 34B was trained on more than 500 billion Arabic tokens.
The 2024 'jais-adapted' models were built on Meta's Llama 2. Jais 2 was trained from scratch.
The JAIS Chat app is published by MBZUAI rather than G42 or Inception, which matters when checking app-store listings.
The four layers
How Jais and ALLaM builds an answer
Every AI answer is made in four layers: what the model learned in pre-training, how it was tuned, what it retrieves live, and how it puts the answer together. Here is each one for Jais and ALLaM, and what it means for a brand that wants to be cited.
Layer 1
Pre-training: the base (foundation) model
Jais 13B learned from 395 billion tokens: 72 billion unique Arabic tokens, repeated to an effective 116 billion, 232 billion English and the rest code, including datasets filtered from Common Crawl. The 2024 Jais family reached 1.6 trillion tokens of public web pages, Wikipedia, news and social content in Arabic and English. Jais 2 was trained from scratch with an Arabic-centric vocabulary. ALLaM 7B learned 4 trillion English tokens, then 1.2 trillion mixed Arabic and English, its Arabic from in-house crawls of the web, news, books, Wikipedia and transcripts. No knowledge cut-offs are published.
What it means for you: This is the layer that decides what these models say about you, because live search is limited. Substantial, well-written Arabic pages that Common Crawl can reach, and that are long and clean enough to pass ALLaM's filters, are how a brand gets learned. Translated fragments and thin landing pages are not.
Layer 2
Post-training: tuning and alignment
Jais-chat was instruction-tuned with safety alignment, and the Jais family's chat models were fine-tuned on a curated mix of Arabic and English instructions. ALLaM 7B Instruct used 7 million instructions and 260,000 preference pairs, and its paper stresses alignment with human preferences. HUMAIN markets HUMAIN Chat as rooted in Arab culture, religion and history, but its alignment is not published.
What it means for you: Tuning to cultural and religious norms means some product categories and claims will be handled cautiously. Content that respects local norms and says plainly what you do is more likely to be repeated than marketing that reads as imported.
Layer 3
Retrieval: live search
HUMAIN Chat has a real-time web search that the user switches on; its provider or index is not published. JAIS Chat documents no web search at all. Neither product documents a crawler, an index or a news partner.
What it means for you: Live retrieval is the smallest lever here and the least documented. Being well indexed in the search engines the region already uses, with Arabic pages that answer questions directly, is the best preparation for whichever provider HUMAIN uses.
Layer 4
The answer: what people read
Whether and how sources are shown is not published for either assistant. HUMAIN Chat's documented features are Arabic voice input across dialects, memory of preferences, switching between Arabic and English, and shareable answer links. Jais 2 runs on Cerebras hardware, which its report says delivers up to about 2,000 tokens a second.
What it means for you: Without visible citations, being named and described correctly is the measure of success. Answers can differ with the language and dialect of the question, so a brand needs testing in Modern Standard Arabic, in the dialects of its markets, and in English.
Crawlers
Jais and ALLaM's crawlers and user agents
The user-agent tokens to look for in your server logs and to name in robots.txt.
| User-agent token | What it does | robots.txt |
|---|---|---|
| Not published Documentation | No crawler or fetcher user agent is documented for Jais, JAIS Chat, ALLaM or HUMAIN Chat. ALLaM's paper says its team crawled Arabic web, news and book data in-house, but gives no user agent or robots.txt token. | Not published |
| CCBot Documentation | Common Crawl's crawler, not run by these developers. Jais 13B's training data included datasets filtered from Common Crawl (Arabic C4 and Pile-CC), and ALLaM's English data included Dolma's Common Crawl portion, so CCBot feeds part of their pre-training indirectly. | Follows robots.txt, Common Crawl says |
HUMAIN's and Inception's websites blocked automated access while we researched this guide, so a crawler page there cannot be ruled out entirely. We will update this guide if one is published.
Crawl sources
Where Jais and ALLaM crawls from, and how to verify it
Not published. G42, Inception, MBZUAI, SDAIA and HUMAIN document no IP ranges, reverse DNS hostnames or networks for any crawler, so there is nothing to verify a request against and no way to attribute visits in your logs to them.
The one route you can verify is indirect. CCBot, whose crawls fed Jais and ALLaM's English data, publishes its IP ranges at index.commoncrawl.org/ccbot.json, and its IPv4 addresses resolve to hostnames under crawl.commoncrawl.org. Our open-weight models guide explains how to check it.
Opting out
How to opt out of Jais and ALLaM, or limit what it uses
There is no assistant-specific token
No crawler token is published for Jais, ALLaM or HUMAIN Chat, so robots.txt cannot target them directly. HUMAIN Chat's live search uses a provider it does not name.
Block Common Crawl to limit future training data
User-agent: CCBot with Disallow: / keeps your pages out of future Common Crawl snapshots and the datasets built from them, which Jais-era models drew on. It does not change past snapshots, existing datasets or released weights, and it affects every model trained on Common Crawl, not just these.
Official pageUse Saudi data protection rights
SDAIA enforces Saudi Arabia's Personal Data Protection Law, in force since 14 September 2023 and fully enforced since 14 September 2024. HUMAIN's own request channel could not be read publicly when we checked.
Accept that released weights stay released
Every Jais generation and ALLaM 7B are downloadable weights. What they learned stays in the copies others have already downloaded.
Official page
Opting out here is mostly indirect and works going forward only. Blocking CCBot also removes you from the data many other models learn from, which can cost you visibility elsewhere. For most brands in the region, the better aim is to be represented accurately, not to be absent.
How we help
How we get brands cited in Jais and ALLaM
The Coaley Peak GEO Research Centre researches and implements GEO for SEO agencies and in-house teams, with foreign markets as our speciality.
Arabic visibility audit
We ask JAIS Chat and HUMAIN Chat what they know about your brand and category, in Modern Standard Arabic, Gulf Arabic and English, and record what is right, wrong or missing.
Arabic content worth learning from
With native Arabic writers, we build substantial pages that pass training-data filters: long enough, clearly Arabic, free of boilerplate, and stating your facts plainly.
Crawlability for the sources these models use
We make sure your Arabic pages are reachable by Common Crawl and indexable by the search engines the region uses, with no client-side rendering hiding content from crawlers.
Base model targeting
Jais and ALLaM 7B are open weights. We probe them directly, with retrieval switched off, to measure what they learned about you, and track the change as new versions are released.
Gulf market entry
We pair Arabic AI visibility with your wider market entry: Google, Bing and the regional assistants, measured together for the UAE and Saudi Arabia.
More on our research, services and team at the Coaley Peak GEO Research Centre.
FAQ
Jais and ALLaM GEO: questions
What is Jais?
Jais is a family of Arabic-first large language models from the UAE, developed by Inception (part of G42), MBZUAI and Cerebras. The first, Jais 13B, was released in August 2023. Jais 2, trained from scratch in 8B and 70B sizes, followed in December 2025 under Apache 2.0 and powers the JAIS Chat app.
What are ALLaM and HUMAIN Chat?
ALLaM is the Arabic and English language model developed by the National Center for AI at SDAIA in Saudi Arabia. HUMAIN, a Public Investment Fund company, calls it its flagship model, and HUMAIN Chat, released in August 2025, runs on ALLaM 34B.
Do Jais or ALLaM crawl my website?
No crawler is documented for either. ALLaM's team says it crawled Arabic web data in-house but publishes no user agent. Jais's training data included datasets filtered from Common Crawl, so Common Crawl's CCBot is the documented, indirect route.
Can I block Jais or HUMAIN Chat in robots.txt?
Not directly, because no token is published. Blocking CCBot limits future Common Crawl-based training data for these and many other models, but it cannot remove anything from weights already released.
Do these assistants show their sources?
Neither documents source citations. HUMAIN Chat has a real-time web search that users can switch on, with a provider it does not name; JAIS Chat documents no web search.
Should we write in Modern Standard Arabic or in dialect?
Both have a place. Modern Standard Arabic suits most product and service pages, while Jais 2 and HUMAIN Chat are tuned to understand dialects too. We test answers in both, because the wording and dialect of a question can change the answer.
Want a free chat about Jais and ALLaM?
Tell us which market and brand you are working on. If we can help, we will set up a free, no-obligation chat with Stephen or Ethan.
More engine guides
Related engines
Independence and trademarks. Jais, JAIS Chat, ALLaM, HUMAIN and HUMAIN Chat are trademarks of their respective owners, including G42, Inception, MBZUAI, SDAIA and HUMAIN or their affiliates. Coaley Peak is independent: we are not associated with, endorsed by or a partner of Inception (G42), MBZUAI, SDAIA and HUMAIN. This guide is our own research from public documentation and reputable reporting, checked in October 2026.
Reviewed by Stephen Grindley, research centre lead. Last checked .
Document reference: ISO_webpage_geo_engine_arabic_ai_assistants_v1
Last modified: 7 October 2026
GEO Research Centre·Engine guides·Jais and ALLaM