Jais 3 Developers Say Closed Models Still Lead on Complex Arabic Agentic Tasks

Jais 3 was developed by the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence and G42's Inception42. Earlier Jais generations focused mainly on understanding and generating Arabic text. The new models add Arabic speech recognition and the ability to follow instructions, use software tools and carry out tasks in sequence. In written answers to Khaleej Times, IFM senior research scientist Haonan Li set out what the models can do and where they fall short.

"We should be clear that progress in Arabic AI does not mean every gap has been closed," Li told the newspaper. "Frontier closed models still lead on complex agentic tasks, such as working through Arabic documents and files, and on depth of cultural knowledge." On one Arabic file-processing evaluation in the developers' own published comparisons, the largest Jais 3 model scored about 64 per cent. A leading closed model scored about 96 per cent.

This matters because the UAE has a national programme to bring agentic AI to 50 per cent of federal government sectors, services and operations within two years. Much of that work runs in Arabic, often mixed with English.

Dialects and Code-Switching

Arabic has more than 400 million speakers. Formal written Arabic can differ sharply from everyday speech in Abu Dhabi, Cairo or Casablanca. Models trained mostly on formal text or on data translated from English can produce fluent Arabic yet misread a regional expression or change dialect without being asked. Research presented this year at the Language Resources and Evaluation Conference tested 19 open models on five dialects (Emirati, Saudi, Egyptian, Syrian and Moroccan) and found large differences in performance between them.

Jais 3 was trained with a focus on dialects, cultural knowledge and the way speakers switch languages. Its companion speech model, Jais Speech, is designed to transcribe Arabic varieties from 19 Arab countries, identify the dialect, and keep English or French terms in their original script when a speaker switches language mid-sentence.

Li gave the example of a government meeting held in Emirati Arabic. Jais 3 could transcribe it, pick out the decisions, assign action items to teams and draft a summary presentation, coordinating the steps through connected tools. "The exact steps completed will depend on the tools and integrations available in the deployment, as well as the level of human review required," Li said.

Six Sizes, Released Openly

Jais 3 comes in six sizes, from 900 million to 375 billion parameters. The smaller versions are meant for modest hardware, possibly including phones. The larger ones target heavier organisational use. Li said the range lets organisations trade off computing cost, performance and where data is processed. A model run locally could also give more control over sensitive information, depending on the safeguards around it.

The models, training checkpoints and evaluation code have been released publicly. IFM and Inception42 say Jais 3 performs strongly against other open models on Arabic understanding and task execution. They also audited existing Arabic benchmarks and published evaluation material for others to check.

Jais 3 now competes with several other Arabic-focused systems, including Saudi Arabia's ALLaM, Qatar's Fanar and the UAE's own Falcon Arabic, as well as international models that keep improving. Li said dialects remain "the hardest frontier for every model" and that reliability is the other unsolved problem across the field. Li argued that progress will depend as much on testing with questions originally written in Arabic, including regional exams and dialect material, as on building larger models.