JAIS Chat

Arabic is spoken by more than 400 million people and is chronically underserved by the models most of the world uses. That is not an oversight — it is what happens when training corpora are assembled from an internet that is overwhelmingly English.

The usual result is an assistant that handles Modern Standard Arabic acceptably, stumbles on the dialects people actually speak, and reasons about culture, law and religion through an English-language lens it never announces.

Jais is the UAE's answer, and JAIS Chat is the consumer-facing front door to it. This review covers what Arabic-first training actually changes, why the dialect problem is harder than it sounds, what the open-weight decision means for organisations, and where a frontier model is still the better tool.

An Arabic Model, Not A Translated One

• JAIS Chat is a bilingual Arabic and English conversational assistant built on Jais, a family of Arabic-first large language models developed in the UAE — the interface is familiar, the model underneath is not.
• Launched by Core42, part of the G42 group, with the underlying models developed by Inception alongside MBZUAI and Cerebras Systems.
• The models were trained Arabic-first rather than adapted: the 13B release used 116 billion Arabic tokens and 279 billion English, and the 30B model used 1.63 trillion tokens in total, of which 475 billion were Arabic.
• Dialect handling is the differentiator. Arabic's written standard differs sharply from how anyone actually speaks it, and models trained mostly on English handle that badly.
• The Jais models are open-weight, which means organisations can evaluate and deploy them independently rather than only through this chat interface.

Model, Provenance And Access

What Arabic-First Training Actually Changes

The distinction between an Arabic-capable model and an Arabic-first one is not marketing, and the token counts make it concrete.

A frontier model trained predominantly on English — Gemini, or any of its peers — will produce grammatical Arabic. What it produces less reliably is Arabic that sounds like it was written by someone who thinks in Arabic — where the idiom is native, the register is right, and the reasoning does not read as a translation of an English thought.

Jais was trained on 116 billion Arabic tokens at the 13B scale and 475 billion at 30B. Those are enormous corpora for a language with a fraction of English's web presence, and assembling them is genuinely difficult work rather than a matter of scraping more.

The English component is deliberate rather than incidental. The 13B model included 279 billion English tokens specifically for cross-lingual transfer — reasoning ability learned in a data-rich language carries across to a data-poor one. Training on Arabic alone would produce a model fluent in Arabic and weak at thinking.

This is why the bilingual framing is honest rather than a compromise. The model is genuinely built for both, with Arabic as the priority rather than the afterthought.

Grammatical Arabic and Arabic that reads as though it was thought in Arabic are different outputs. Only the second one requires the training corpus to have been built deliberately.

The Dialect Problem Is Larger Than Outsiders Assume

Arabic has a structural feature that makes it unusually hard for language models, and it is worth explaining because it is the crux of the whole product.

Modern Standard Arabic is the formal written language: news, literature, official documents, education. It is broadly shared across the Arab world and it is what almost all written training data consists of.

It is also not what anyone speaks. Egyptian, Levantine, Gulf, Maghrebi and Iraqi Arabic differ from Modern Standard and from each other in vocabulary, grammar and pronunciation — in some pairings enough to impede mutual comprehension.

So a model trained on written Arabic learns the version of the language people write and not the version they speak, which is a serious problem for a conversational assistant. Someone typing casually in Gulf Arabic is not writing Modern Standard, and a model that only knows the formal register will misread the question or answer in a way that lands as stiff and foreign.

JAIS Chat's stated handling of multiple dialects and cultural nuance is the correct thing to target. It is also the thing you should test hardest, in your own dialect, because it is where the difference between this and a general model will show up first if it exists at all.

• Modern Standard Arabic dominates written corpora but is spoken by nobody conversationally
• Egyptian, Levantine, Gulf, Maghrebi and Iraqi Arabic differ substantially from it and from each other
• Casual typed Arabic mixes dialect, Modern Standard and sometimes Latin transliteration
• Cultural, religious and legal context varies across the region and rarely appears in English-trained corpora
• Right-to-left rendering and Arabic script handling are a separate class of engineering problem again

Open Weights Change Who This Is For

The Jais models were released openly, and that decision does more for the region than the chat interface itself.

An open-weight model can be downloaded, evaluated on your own benchmarks, fine-tuned on your own data, and deployed inside your own infrastructure. For a government department, a bank or a hospital in the region, that is the difference between an option and a non-starter.

Data residency is usually the blocking constraint. Regulated organisations across the Gulf are frequently prohibited from sending citizen or customer data to a model hosted elsewhere, and no amount of quality closes that gap. A model you can run yourself does.

It also means the models can be independently evaluated, which matters for a claim like best-performing Arabic model. Open weights turn that from an assertion into something researchers can test.

The practical implication for a technical reader: JAIS Chat is the demonstration, not the deployment. If you are building an Arabic product, the interesting artefact is the model on Hugging Face rather than the chat window.

Note also that a later generation exists. Jais 2 is a 70 billion parameter model from the same Inception, Cerebras and MBZUAI partnership — so check which model any given interface is actually serving before benchmarking against it.

Where A Frontier Model Still Wins

An honest assessment has to be specific about the gap, and it is a real one.

The largest models from OpenAI, Anthropic and Google are trained at a scale Jais does not match, and that shows in general reasoning, coding, mathematics, long-context work and tool use. If your task is hard reasoning that happens to be phrased in Arabic, a frontier model will often do better even without Arabic-first training.

The surrounding product also differs enormously. ChatGPT and Claude arrive with file handling, web browsing, code execution, mobile apps, developer platforms and integrations built over several years. A newer regional assistant is competing on the model, not on the ecosystem.

Where Jais wins is narrower and real: dialectal Arabic, culturally grounded responses, Arabic-language content generation that does not read as translated, and deployment inside jurisdictions where the alternative is not permitted to run at all.

The sensible position for most Arabic-speaking users is not to choose. Use a frontier model for reasoning-heavy work and an Arabic-first model where the language and the cultural context are the point.

1. Write the same prompt in your own dialect and in Modern Standard, and compare how each model handles both.
2. Ask for something culturally specific — a regional holiday, a local legal norm, a religious question — and check the framing, not just the facts.
3. Ask for generated Arabic content and have a native reader judge whether it reads as translated.
4. Run a reasoning-heavy task in Arabic and compare against a frontier model honestly.
5. If you are deploying rather than chatting, evaluate the open weights directly rather than the chat interface.

What To Confirm Before Organisational Use

Two practical points, and one clarification about names that trips people up.

Confirm which entity you are dealing with. The Jais models and the Core42 launch are well documented in the record. If you are procuring anything, verify the operator, the terms and the data handling of the specific interface you are using — the model being open means more than one service can serve it.

Check data handling for your jurisdiction. The whole appeal of a regional model is usually regional data handling, so confirm where conversations are processed and retained rather than assuming that follows from where the model was built.

Distinguish the model from the product. Jais is the model family. JAIS Chat is one interface onto it. Jais 2 is a later, larger generation. Benchmarks published about one do not automatically describe another, and mixing them up is the most common error in coverage of this project.