MBZUAI and Inception42 unveil Jais 3, a comprehensive suite of Arabic artificial intelligence models
The latest generation of the Jais family brings sophisticated language understanding and speech recognition to Arabic across dialects, from Modern Standard Arabic to regional variations spoken across the Arab world.
Mohamed bin Zayed University of Artificial Intelligence, through its Institute of Foundation Models, and Inception42 have unveiled Jais 3, an expansive collection of artificial intelligence models engineered to serve the Arabic language with exceptional sophistication and cultural depth. This latest iteration represents a significant advancement for the Jais family, which has established itself as a cornerstone of Arabic artificial intelligence development since its inaugural launch in Abu Dhabi.
Jais 3 encompasses a carefully calibrated fleet of six open language models, ranging from 900 million to 375 billion parameters, working in concert with Jais Speech, a specialised speech recognition model designed to comprehend Arabic and its diverse dialects. This comprehensive approach ensures that the technology can engage meaningfully with both written and spoken Arabic, transcribing conversations, generating text, and executing practical workflows including document creation and presentation assembly.
The technological foundation upon which Jais 3 rests is K2 Horizon, the Institute of Foundation Models’ fully open-source foundation model fleet released in September 2026. The models have been trained extensively on Modern Standard Arabic alongside regional dialects spanning from Emirati and Gulf variations to North African Arabic, incorporating the cultural knowledge and linguistic nuances embedded within a language spoken by more than 400 million people globally.
Jais Speech represents a particularly noteworthy achievement, transcribing Modern Standard Arabic and dialects across 19 Arab countries under real-world conditions. The system gracefully follows speakers as they transition between Arabic and languages such as English or French, preserving those multilingual utterances in their original scripts whilst identifying the dialect being spoken and applying Arabic vowel marks to transcripts when desired. This sophistication reflects a profound understanding that Arabic communication, as it exists authentically, frequently encompasses code-switching and dialectal variation.
Independent evaluation across an audited suite of 24 Arabic benchmarks has demonstrated the models’ considerable capabilities. The largest Jais 3 model leads comparable open models on agentic tasks and achieves parity with the strongest alternatives on Arabic comprehension. Jais Speech recorded the lowest average word and character error rates amongst all models evaluated, whilst the 7-billion and 32-billion-parameter variants achieved the highest average Arabic scores amongst open models of comparable scales.
The introduction of Jais 3 reflects the United Arab Emirates’ sustained commitment to artificial intelligence advancement and aligns with the National AI Strategy 2031. By releasing these models openly alongside comprehensive benchmarks, MBZUAI and Inception42 are advancing Arabic language capabilities for researchers, developers and organisations across the Arabic-speaking world, ensuring that artificial intelligence remains accessible, practical and culturally attuned to the communities it serves.
Source: WAM