The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42 introduced Jais 3 on October 15, broadening their Arabic AI offering across speech, writing and tool-connected tasks.
The release includes six language models ranging from 900 million to 375 billion parameters, alongside Jais Speech. The developers say the models are available through Hugging Face.
Everyday language is part of the task
Jais Speech targets Modern Standard Arabic and regional dialects, including conversations in which speakers switch between Arabic, English and French.
The language models also target Arabic document work. Connected to software tools, the developers say they can research topics and produce or edit documents, spreadsheets and presentations. A separate agentic platform is described as coming soon.
Performance comparisons in the announcement come from IFM and Inception42’s evaluations. They provide evidence under the developers’ test conditions rather than independent confirmation of leadership across Arabic applications.
A correct transcript is only the beginning
For a workplace assistant, recognizing words is one stage of a longer process. The system must also identify what a speaker wants, preserve names and numbers, and produce material in the appropriate register.
Consider an assistant turning a spoken request into a formal appointment letter. It would need to distinguish conversational phrasing from the wording suitable for the document, retain the appointment details and avoid adding information that was never supplied. This illustrates an evaluation scenario, rather than a reported deployment.
Similar checks apply when a model can use tools. Producing a plausible plan does not establish that each action was completed correctly.
Testing the complete workflow
AIYXL Perspective: Language-specific development becomes valuable when it improves the work people can complete. For Arabic AI, useful testing should include dialect variation, mixed-language input, document consistency and recovery from misunderstandings.
Organizations assessing Jais 3 can begin with representative tasks reviewed by local users. Measuring how much correction remains—from the first spoken instruction to the final document—would offer a practical view of its value.