MBZUAI and Inception42 Launch Jais 3 Open Arabic AI Fleet Spanning 900M to 375B Parameters and Multilingual Speech Engine

ABU DHABI — The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), operating through its Institute of Foundation Models (IFM), in collaboration with Inception42, has officially launched Jais 3, a comprehensive fleet of open source artificial intelligence models engineered natively for Arabic language comprehension, generation, speech recognition, and autonomous agentic workflows.

The release marks a substantial evolution of the sovereign Jais franchise first pioneered in Abu Dhabi, bridging linguistic fluency across more than 400 million native speakers. The Jais 3 fleet integrates six open language models ranging in architecture from 900 million up to 375 billion parameters, paired with Jais Speech, an advanced acoustic speech recognition system. All architectures in the collection are built upon K2 Horizon, the fully open source foundational architecture suite released by the IFM in September 2026.

Engineered to reflect spoken and written linguistic nuances, Jais 3 is trained across Modern Standard Arabic alongside diverse regional dialects spanning the Arabian Gulf, the Levant, and North Africa. The architecture accounts for regional colloquialisms and real time code switching, dynamically following speakers who alternate between Arabic, English, or French mid conversation while maintaining accurate orthography and script integrity.

In standardized performance evaluations conducted across an audited suite of twenty four public Arabic benchmarks, the top tier 375 billion parameter Jais 3 model secured leadership among open models in complex agentic execution while matching peak frontier baselines in semantic Arabic comprehension. Concurrently, the small and mid size configurations, including the 7 billion and 32 billion parameter models, recorded the highest average Arabic benchmark scores among open models in comparable parameter classes. Jais Speech recorded the lowest average word and character error rates across 19 Arab countries, demonstrating the capability to apply automatic Arabic diacritics and vowel marks directly to live audio transcripts.

Richard Morton, Vice President and Executive Director of the IFM, stated that Jais 3 demonstrates the tangible output of the UAE’s sustained sovereign investments in open artificial intelligence research, noting that expanding the family into voice and agent capabilities directly advances the mandates of the UAE National AI Strategy 2031.

Ashish Koshy, Chief Executive Officer of Inception42, underscored that developing an Arabic foundation model transcends simple data translation, requiring native training on colloquial vernaculars, syntactic variations, and deep cultural contexts. Koshy noted that uniting six distinct model weights enables Jais 3 to lead Arabic benchmarks across every operational size class.

Dr. Larry Murray, Senior Vice President of Applied Science at Inception42, highlighted that reinforced reasoning, instruction following, and native tool calling capabilities establish Jais 3 as an operational engine for multi step agentic enterprise workflows, allowing commercial enterprises and public sector agencies to align computational requirements with specific infrastructure budgets.