Abu Dhabi Unveils Jais 3 AI to Understand Arabic Dialects and Multilingual Speech
Jais 3, Abu Dhabi's new AI model, recognizes Arabic dialects, retains English terms, and completes complex tasks from speech.
Abu Dhabi has introduced Jais 3, an advanced Arabic artificial intelligence system designed to comprehend various Arabic dialects and seamlessly integrate English technical terms within conversations. Developed by Mohamed bin Zayed University of Artificial Intelligence's Institute of Foundation Models (IFM) and G42's Inception42, Jais 3 was launched in October as part of a new generation of AI models focused on practical language understanding and task execution.
Unlike earlier versions that primarily focused on language recognition and generation, Jais 3 combines Arabic speech recognition with capabilities to follow instructions, utilize software tools, and perform sequences of tasks. This advancement aims to allow Arabic speakers to communicate naturally without altering their dialect or mixing languages for the AI to understand them effectively.
Haonan Li, a senior research scientist at IFM, highlighted the importance of this approach, noting that Arabic speakers often switch between dialects, Modern Standard Arabic, and other languages such as English within the same conversation. Jais 3 is designed to interpret these shifts contextually rather than treating them as separate issues.
Arabic, spoken by over 400 million people, varies widely across regions, making it challenging for AI systems trained mainly on formal written Arabic or translated datasets to accurately interpret everyday speech. Research presented at the Language Resources and Evaluation Conference revealed significant performance disparities among AI models when processing different Arabic dialects, underscoring the need for dialect-sensitive training.
Jais 3 addresses these challenges by focusing on dialectal variations, cultural context, and code-switching between Arabic and other languages. Its speech model, Jais Speech, can transcribe dialects from 19 Arab countries, identify specific dialects, and preserve English or French terms in their original script during transcription.
This capability is particularly beneficial for UAE workplaces where Arabic and English coexist, enabling more practical use of AI tools. Beyond transcription, Jais 3 can also execute complex tasks. For example, after a government meeting conducted in Emirati Arabic, the AI can transcribe the discussion, extract decisions, assign tasks to relevant teams, and generate a presentation summarizing the outcomes—all through integrated software tools.
Li described this as a shift towards agentic AI, which moves beyond answering questions to performing a series of connected actions. The specific tasks completed depend on the available software integrations and the degree of human oversight.
Jais 3 is available in six model sizes, ranging from 900 million to 375 billion parameters, allowing deployment on devices from smartphones to large organizational systems. This variety offers organizations flexibility in balancing computing resources, performance, and data privacy, especially when models are deployed locally.
The models, training data checkpoints, and evaluation code have been made publicly accessible, enabling developers to customize the technology and independently assess its performance. This openness supports the UAE’s national initiative to implement agentic AI in 50 percent of federal government sectors within two years.
While Jais 3 shows strong performance compared to other open Arabic AI models, including Saudi Arabia’s ALLaM, Qatar’s Fanar, and the UAE’s Falcon Arabic, some proprietary models still outperform it on complex tasks. For instance, in an Arabic file-processing evaluation, the largest Jais 3 model scored around 64 percent, whereas a leading closed-source model achieved approximately 96 percent.
Li acknowledged these gaps, emphasizing that although progress has been made, frontier closed models continue to lead in handling complex agentic tasks and possess deeper cultural knowledge.
What this means for you
- Jais 3 AI can understand Arabic dialects and English terms used in conversations, improving communication with AI tools.
- The AI can transcribe meetings in Emirati Arabic and assist in assigning tasks and preparing summaries automatically.
- Different sizes of the AI model allow use on devices ranging from smartphones to organizational systems, offering flexibility.
- The UAE plans to integrate agentic AI like Jais 3 across half of federal government sectors within two years.