Jais 3 integrates 19 dialects into open models, a new platform that tackles Modern Standard Arabic rebound and Gulf data scarcity
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi, in partnership with Inception 42, launched the Jais 3 platform, an integrated Arabic AI system that aims to transcribe spoken conversations, draft documents, and create presentations across multiple dialects. The platform includes open-source language models in two main sizes, seven billion and thirty-two billion parameters, and the university says it achieved the highest average scores on Arabic language tests among open models of comparable size. It also includes a speech component called Jais Speech, which recognizes Modern Standard Arabic as well as dialects from ten Arab countries, and tracks speakers when they switch abruptly between Arabic and English or French mid-sentence and then return, registering the lowest word- and character-error rates among all compared models.
The biggest engineering challenge in training the models was not gathering general text, but breaking their automatic tendency to revert to Modern Standard Arabic and overcoming the scarcity of local dialect data.Hwanan Lee, senior research scientist at the Institute for Foundational Models at the university, explains that the team had to work with a very limited open-source dataset for Gulf dialects, as well as correct a recurring algorithmic behavior that forces models to drift back to Modern Standard Arabic and ignore the target dialect. The same problem occurred in Morocco, where the available data were heavily skewed toward Moroccan Darija at the expense of other local dialects, compelling developers to resort to synthetic data generation and cleaning to fill gaps and provide a balanced representation.
The speech component provides advanced processing beyond simple audio-to-text conversion; it preserves words borrowed from English or French in their original Latin script rather than phoneticizing them into distorted Arabic letters, automatically classifies the spoken dialect, and adjusts the extracted text with appropriate diacritics. The university and its partner subjected the platform to an evaluator that uses twenty-four Arabic benchmark metrics released with the models for testing, while Richard Morton of the Institute for Foundational Models linked the effort to the UAE National Artificial Intelligence Strategy 2031, and Dr. Larry Murray of Inception 42 described the models as a cornerstone for building self-directed workflows for intelligent agents that organisations can align with their available compute budgets.
As editor-in-chief monitoring this industry through the lens of market integrity, I consider performance-figure auditing a professional prerequisite before succumbing to promotional hype; the data on reduced word errors and superior benchmark scores were released by the model developers themselves and have not yet undergone independent, neutral third-party review in real-world operating environments. Relying on laboratory results without field testing exposes companies to operational risks, especially since government strategy statements and intelligent-agent descriptions remain marketing narratives from sponsoring firms until their actual effectiveness is demonstrated in contact centers and complex operating systems.
If you lead a technical team in the Gulf, Egypt, or the Levant, this development gives you an operational option to bypass the consumption of ready-made global models by hosting open weights locally and building voice-transcription tools for customer service without paying recurring per-call fees in dollars to Western vendors. This approach requires you to allocate budgets for suitable computing infrastructure and attract engineers capable of integrating dialect-aware speech systems into your business applications, while ensuring that model outputs undergo stringent quality testing matched to your local audience’s dialect before final integration into production pipelines.