MBZUAI and Inception42 release Jais 3 Arabic language models

MBZUAI and Inception42 released Jais 3, a suite of six open-source Arabic language models covering 19 dialects. The models range from mobile to flagship scale.

Six models from mobile to flagship#

The Jais 3 suite comprises six models designed for different deployment scenarios. The range extends from lightweight mobile implementations to flagship-scale systems. Each model in the suite is released under an open-source licence, allowing developers to access and deploy the technology without proprietary restrictions. The models support 19 Arabic dialects, addressing the linguistic diversity across Arabic-speaking regions. MBZUAI developed the suite in collaboration with Inception42.

Open-source release for Arabic language processing#

The models are available as open-source releases, removing licensing barriers for developers working on Arabic language applications. The suite addresses the technical challenge of building language models that handle multiple Arabic dialects within a single framework. The release includes models optimised for mobile devices, where computational resources and memory are constrained, alongside larger models designed for server deployment. The 19 dialects covered span the geographic range of Arabic usage.

MBZUAI and Inception42 collaboration#

MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence, partnered with Inception42 to develop and release the Jais 3 suite. The collaboration combined MBZUAI's research capabilities with Inception42's development resources. The partnership focused on creating models that could be deployed across different hardware configurations, from mobile phones to data centre infrastructure. The open-source approach allows third parties to build applications on the models without negotiating commercial terms.

Dialect coverage and deployment options#

The 19 dialects supported by Jais 3 include regional variations of Arabic used across the Middle East and North Africa. The suite's architecture allows a single model to process multiple dialects rather than requiring separate models for each variant. The mobile-optimised models are designed to run on devices with limited processing power and storage, while the flagship models target use cases requiring higher accuracy and broader language understanding. The release provides developers with a choice of model size based on their deployment constraints and performance requirements.