The new family of Arabic artificial intelligence models, Jais 3, was unveiled in Abu Dhabi in October. These models were developed by the Institute for Fundamental Models (IFM) at Mohammed bin Zayed University of Artificial Intelligence and Inception42 from G42. This technology marks a shift in the UAE's approach to developing Arabic AI.
Previous versions of Jais primarily focused on language generation and understanding. However, the new models combine Arabic speech recognition with the ability to follow instructions, utilize software tools, and execute sequences of tasks.
According to Haonan Li, a senior researcher at IFM, this distinction is critically important because native Arabic speakers should not have to change their way of communicating just so that the technology can understand them. Li noted that the Arabic language varies significantly across different regions, and people freely switch between dialects, Modern Standard Arabic, and other languages.
He gives an example where a person might speak Emirati Arabic, switch to English to use a technical term, and then return to Arabic. The model must be able to understand this complex switching in context, rather than treating each such switch as a problem.
Arabic is spoken by over 400 million people, but it is far from monolithic. Formal Arabic, used in official documents and newspapers, can differ greatly from everyday speech in cities like Abu Dhabi, Cairo, or Casablanca. This creates difficulties for Arabic AI systems trained predominantly on formal written materials or data translated from English.
Research published this year at the Conference on Language Resources and Evaluation analyzed 19 open AI models across five Arabic dialects, including Emirati, Saudi, Egyptian, Syrian, and Moroccan. Scientists found significant performance differences between dialects, highlighting how far this field is from achieving stable comprehension.
Jais 3 aims to solve this problem through training focused on Arabic dialects, cultural knowledge, and language-switching methods. Its companion speech model, Jais Speech, is designed to transcribe various variants of Arabic used in 19 Arab countries. It is capable of identifying dialects and retaining English or French terminology in its original spelling when the speaker changes languages.
For residents and organizations in the UAE, this could make AI tools more useful in workplaces where both Arabic and English are used simultaneously. However, accurate transcription is only part of the development. Haonan Li provided an example where a civil servant concludes a meeting conducted in Emirati Arabic. The employee can instruct Jais 3 to transcribe the conversation, identify decisions made, assign tasks to relevant teams, and prepare a presentation summarizing the results.
The system can coordinate work through connected software tools instead of processing each stage separately. Li explained that this demonstrates the direction of agentic AI: moving from simply answering a question to executing an entire sequence of tasks. He added that the precise steps depend on available tools and integrations during implementation, as well as the required level of human verification.
Jais 3 is available in six sizes, ranging from 900 million to 375 billion parameters, which are numerical values learned by the model during training. Smaller versions are intended for more modest hardware, potentially including mobile phones, while larger systems target demanding organizational applications. Li noted that this range allows organizations to choose between computational cost, performance, and data processing location. A locally deployed model can also provide greater control over sensitive data, depending on security measures.
The models, their training checkpoints, and evaluation code have been made public, allowing developers to adapt the technology and independently verify its effectiveness. This approach is particularly relevant given the UAE's national plan to implement agentic AI in 50 percent of federal government sectors, services, and operations within two years.
Competition is no longer limited to one Emirati model. Several Arabic AI systems exist in the region, including ALLaM from Saudi Arabia, Fanar from Qatar, and Falcon Arabic from the UAE, alongside increasingly capable international models. IFM and Inception42 report that Jais 3 shows strong results compared to competing open models in Arabic language understanding and task execution. Researchers also audited existing Arabic benchmarks and published evaluation materials for others to study.
Nevertheless, their published comparisons show that some leading proprietary models still outperform Jais 3 in complex tasks. In one test involving Arabic file processing, the largest Jais 3 model scored about 64 percent, compared to approximately 96 percent for a leading closed model. Li acknowledged that the remaining gaps are significant. He emphasized that progress in Arabic AI does not mean all shortcomings have been eliminated; advanced closed models still lead in complex agentic tasks, such as working with Arabic documents and files, as well as in depth of cultural knowledge. Dialects remain the most challenging frontier for any model, and reliability is another open issue in this field.
For researchers, this raises the question: does the model truly understand Arabic, or has it merely learned to reproduce familiar linguistic patterns? Li believes that assessing this difference requires questions originally written in Arabic, including regional and dialectal materials, rather than relying solely on translated tests in English. Practical implications go beyond benchmarks. An AI assistant that occasionally misunderstands a dialect can be inconvenient in casual conversation. Such misunderstanding becomes more serious when it affects a government document, medical record, or business instruction. Therefore, according to Li, progress will depend as much on improving reliability and evaluation as on creating larger models. The goal is not just to create more Arabic models, but to develop systems that reliably work for Arabic speakers in the real world. The next stage of Arabic AI development, therefore, may be judged not so much by how naturally the machine can hold a conversation, but by whether people can trust it to understand their instructions and perform the work.
