Abu Dhabi, UAE – The UAE has unveiled Jais 3, the latest generation of its Arabic-first large language model, alongside Jais Speech, a new voice recognition model. Both were developed by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) together with Abu Dhabi-based Inception42.
The new models are designed to work the way Arabic speakers actually talk. Jais 3 covers Modern Standard Arabic as well as dialects from 19 Arab countries, including Gulf and North African varieties, serving a language spoken by more than 400 million people.
Built for real conversations
Beyond answering questions, Jais 3 can carry out practical tasks such as transcribing conversations and preparing documents and presentations. It accepts spoken instructions in Arabic or English and acts on them.
- Code-switching: handles speakers who mix Arabic with English or French in the same sentence, keeping foreign words in their original script.
- Dialect detection: identifies which dialect is being spoken.
- Diacritics: can add Arabic vowel marks (tashkeel) to transcripts for clearer reading.
Benchmark performance
According to the developers, the 7-billion and 32-billion-parameter versions of Jais 3 recorded the highest average Arabic scores among open models of comparable size, while the largest model leads open models on agentic tasks. Jais Speech achieved the lowest average word and character error rates among the models it was compared with.
The evaluation used 24 Arabic benchmarks, which are being released publicly so that other researchers can test and compare their own models.
Why it matters for UAE businesses
Most mainstream AI tools are trained primarily on English and struggle with regional dialects. Models that understand Emirati, Egyptian, Levantine or Maghrebi Arabic open the door to better customer service bots, transcription for government and healthcare services, and Arabic content tools for businesses across the region.
Source: The National.
