A government meeting conducted in Emirati Arabic may move naturally between local expressions, formal language and English technical terms. The decisions made during that meeting must eventually become written records, assigned tasks and perhaps a presentation. For an artificial intelligence system, understanding the words is only the beginning. The harder challenge is turning that conversation into completed work without losing its meaning.
That is the ambition behind Jais 3, a new family of Arabic AI models launched in Abu Dhabi in October by Mohamed bin Zayed University of Artificial Intelligence’s Institute of Foundation Models (IFM) and G42’s Inception42.
The technology represents a shift in the UAE’s approach to Arabic AI. Earlier generations of Jais concentrated largely on understanding and generating language. The new models combine Arabic speech recognition with the ability to follow instructions, use software tools and perform sequences of tasks.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
For Haonan Li, senior research scientist at IFM, the distinction matters because Arabic speakers should not have to change how they communicate simply to make technology understand them. “Arabic varies significantly across regions, and people move naturally between dialects, Modern Standard Arabic and other languages,” Li told Khaleej Times in written responses.
“Someone might speak in Emirati Arabic, switch into English for a technical term, and then return to Arabic. A model needs to understand that entire interaction in context, rather than treating each shift as a problem.”
The challenge of everyday Arabic
Arabic is spoken by more than 400 million people, but it is far from uniform. The formal Arabic used in newspapers and official documents can differ considerably from everyday conversation in Abu Dhabi, Cairo or Casablanca.
That creates a problem for Arabic AI systems trained primarily on formal written material or datasets translated from English. A system may produce convincing Arabic while misunderstanding a regional expression, switching dialects unexpectedly or missing the intended meaning of an instruction.
Research published this year in the Language Resources and Evaluation Conference examined 19 open AI models across five Arabic dialects, including Emirati, Saudi, Egyptian, Syrian and Moroccan. The researchers found substantial differences in performance between dialects, highlighting how far the field remains from consistent understanding.
Jais 3 attempts to address that problem through training focused on Arabic dialects, cultural knowledge and the way people switch between languages. Its accompanying speech model, Jais Speech, is designed to transcribe Arabic varieties spoken across 19 Arab countries. It can identify dialects and preserve English or French terminology in its original script when a speaker changes languages.
For residents and organisations in the UAE, that could make AI tools more practical in workplaces where Arabic and English are used together. But accurate transcription is only part of the development.
From answering questions to completing tasks
Li offered the example of a government employee finishing a meeting conducted in Emirati Arabic. The employee could instruct Jais 3 to transcribe the conversation, identify decisions, assign action items to the relevant teams and prepare a presentation summarising the outcomes.
Instead of handling each stage separately, the system could coordinate the work through connected software tools.
“This illustrates the direction of agentic AI: moving from answering a question to carrying out a sequence of tasks,” Li explained. “The exact steps completed will depend on the tools and integrations available in the deployment, as well as the level of human review required.”
Jais 3 comes in six sizes, ranging from 900 million to 375 billion parameters, the numerical values a model learns during training. Smaller versions are designed for more modest hardware, potentially including phones, while larger systems target demanding organisational applications.
Li said this range gives organisations choices over computing costs, performance and where information is processed. A locally deployed model could also offer greater control over sensitive data, depending on the safeguards surrounding it.
The models, their training checkpoints and evaluation code have been released publicly, allowing developers to adapt the technology and independently examine its performance.
That approach is particularly relevant as the UAE pursues a national programme to introduce agentic AI across 50 per cent of federal government sectors, services and operations within two years.
Gap between fluency and reliability
The competition is no longer confined to one Emirati model.
The region now has several Arabic-focused AI systems, including Saudi Arabia’s ALLaM, Qatar’s Fanar and the UAE’s Falcon Arabic, alongside increasingly capable international models.
IFM and Inception42 report that Jais 3 performs strongly against competing open models across Arabic understanding and task execution. The researchers also audited existing Arabic benchmarks, and released evaluation material for others to examine.
Yet their published comparisons show that some leading proprietary models still outperform Jais 3 on demanding tasks. On one Arabic file-processing evaluation, the largest Jais 3 model scored about 64 per cent, compared with approximately 96 per cent for a leading closed model.
Li acknowledged that the remaining gaps are substantial.
“We should be clear that progress in Arabic AI does not mean every gap has been closed. Frontier closed models still lead on complex agentic tasks, such as working through Arabic documents and files, and on depth of cultural knowledge. Dialects remain the hardest frontier for every model. Reliability is the other open problem across the field.”
For researchers, this raises another question: does a model genuinely understand Arabic, or has it simply become skilled at reproducing familiar language patterns?
Li said evaluating that distinction requires questions originally written in Arabic, including regional examinations and dialectal material, rather than relying entirely on translated English tests.
The practical consequences extend beyond benchmarks. An AI assistant that occasionally misunderstands a dialect may be inconvenient in casual conversation. The same misunderstanding becomes more consequential when it affects a government document, medical record or business instruction.
For Li, progress will therefore depend as much on improving reliability and evaluation as on building larger models.
“The objective is not simply to build more Arabic models, but to develop systems that work reliably for Arabic speakers in the real world.”
The next stage of Arabic AI may consequently be judged less by how naturally a machine can hold a conversation than by whether people can trust it to understand their instructions and finish the work.
Source: Khaleej Times
