Arabic LLM UAE: what Jais 3 means for companies building AI agents
Zeta42 · 16 October 2026
MBZUAI and Inception42 released Jais 3, six open Arabic language models and a speech model. Here is what a UAE company should test before building on them.
A new Arabic LLM for UAE companies to test arrived on 15 October. According to Gulf News, MBZUAI's Institute of Foundation Models (IFM) and Inception42 released Jais 3, a set of six open language models built for Arabic, together with Jais Speech, a speech recognition model for Arabic and its dialects. For a company planning an AI agent or a chat or voice interface that has to work in Arabic, this is a fresh set of models to put through a trial.
What was released
The six language models range from 900 million to 375 billion parameters. All are built on K2 Horizon, IFM's open-source foundation model fleet released in September 2026, and all share a context window of 512,000 tokens, which the article describes as the longest of any Arabic-native model. They are trained on Modern Standard Arabic and on regional dialects, including Emirati, Gulf and North African varieties. Each model ships with five training checkpoints, and the models and the speech model are published on a public model hub. Jais 3 was developed with Cerebras Systems.
Why an Arabic LLM in the UAE has to handle dialect
Staff and customers in the UAE do not write or speak only in formal Arabic. They switch between dialects, and many switch into English or French mid-sentence. Jais Speech is built for exactly that: it transcribes Arabic dialects, follows speakers who change language mid-sentence, identifies the dialect and can add vowel marks. IFM and Inception42 report that it had the lowest word and character error rates among the models they compared, which included commercial services.
What the agent results say, and what they do not
IFM and Inception42 tested the models on 24 Arabic benchmarks and are releasing that suite publicly. They report that the largest model leads open models on agentic tasks. At 7 billion parameters, the average Arabic evaluation score rose from 56.4 for the K2 Horizon base model to 79.6 for Jais 3. These are results published by the people who built the models. They show the models are worth testing, and they say little about how a model will behave on your own contracts, purchase orders or HR letters.
Where the long context window helps
A 512,000-token window means a model can be given a very long document, or a set of documents, in one go. For a company that keeps contracts, policies and correspondence in Arabic, that matters more than any single benchmark score. Whether the models stay accurate across that much text is something to measure with your own files.
What to test before you build
Start with a small sample of real documents and real requests, in the Arabic your staff actually use. Check dialect handling, mixed Arabic and English text, and long documents. Then decide who reviews the cases the model gets wrong, because that decision shapes the whole design.
This is the work Zeta42 does when it builds AI agents. Our agents read documents and act inside an operation, for procurement, HR and finance. They work within limits the client sets, every action is logged, and they ask a person only when the documents do not agree. The same approach applies to chat and voice interfaces built on a client's own data, where an Arabic speech model decides how well a caller is understood.
The practical next step
Pick one process that runs on Arabic documents, such as supplier invoices or leave requests, and test two or three models against twenty real examples. Record where each one fails, and compare that with the cost of a person checking the exceptions. If the failures are rare and easy to catch, you have a case for building. If they are frequent, you have learned that cheaply, before any system exists.
Common questions
- What is Jais 3?
- Jais 3 is a set of six open Arabic-focused language models from MBZUAI's Institute of Foundation Models and Inception42, ranging from 900 million to 375 billion parameters. It was released alongside Jais Speech, a speech recognition model for Arabic and its dialects.
- Can a UAE company host an Arabic model on its own network?
- The models are released as open models, so hosting one yourself is an option to assess. Zeta42 runs AI models on-premise for data residency, and says plainly when the cloud is fine, because on-premise costs more up front and earns it back only if the constraint is real.
- How should a company test an Arabic LLM before building an agent?
- Run a small sample of your own documents and requests through each model, in the dialects and mixed languages your staff use. Record the failures and compare them with the cost of a person reviewing the exceptions.
Source: gulfnews.com