The Mohamed bin Zayed University of Artificial Intelligence through its Institute of Foundational Models collaborated with Inception42 to release the Jais 3 collection of models on Thursday, the Emirates News Agency reported. The suite consists of six open language models with parameter sizes from 900 million to 375 billion in addition to an audio model built for Arabic speech recognition that handles multiple dialects. These systems excel at comprehending both spoken and written Arabic across variations while generating coherent text and executing practical assignments such as transcribing dialogues or creating documents and presentations.
A joint evaluation conducted by the Institute of Foundational Models and Inception42 tested the new models against 24 distinct Arabic performance standards now made publicly available. The assessment found that the largest model surpassed other open systems in agent-related functions and matched leading alternatives in overall Arabic language understanding. The audio component recorded the lowest average error rates for both words and characters when compared with all evaluated models.
Models sized at 7 billion and 32 billion parameters achieved the highest average results on Arabic language examinations among open equivalents of matching scale according to the developers’ data. Smaller and midsize variants in the Jais 3 lineup outperformed comparable open models across the tested criteria the evaluation showed. Such outcomes build directly on the original Jais project first introduced in 2023 by the same partners alongside Cerebras Systems which at the time established itself as the top open Arabic large language model per industry analyses from that period.
Subsequent iterations including Jais 2 released in late 2025 expanded training to trillions of carefully curated tokens spanning Arabic English and code as documented in technical reviews from sources such as Khaleej Times. Those earlier systems demonstrated strong bilingual performance and rapid inference speeds when deployed on specialized hardware. The latest Jais 3 advances this trajectory by integrating voice command capabilities that accept instructions in either Arabic or English and execute them seamlessly through the combined language and audio models.
The developments reflect ongoing efforts by UAE entities to strengthen sovereign capabilities in Arabic artificial intelligence where demand continues to grow for culturally attuned systems. Inception42 has maintained a consistent role in these initiatives evolving from foundational research toward deployable solutions that support government and enterprise applications. Public leaderboards hosted on platforms like Hugging Face have repeatedly highlighted the Jais family’s advantages in areas such as instruction following and generative accuracy for Arabic content.
Integration of the audio model with the language variants enables end-to-end voice-driven workflows that maintain linguistic consistency even when switching between dialects or formality levels the announcement detailed. Tasks ranging from summarization to complex document assembly benefit from the models’ ability to follow detailed directives with precision. This launch provides developers with openly accessible tools that address previous gaps in high-quality Arabic AI resources while aligning with broader regional priorities for technological self-reliance.
