Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), its Institute of Foundation Models (IFM) and G42’s Inception42 have launched Jais 3, a family of six open Arabic language models and Jais Speech, an Arabic speech recognition model.
Developed with Cerebras Systems, the models range from 900 million to 375 billion parameters and are built on IFM’s open-source K2 Horizon foundation model fleet. They support Modern Standard Arabic, regional dialects including Emirati, Gulf and North African Arabic, and multi-step agentic tasks. All six models use the same training approach, support a 512,000-token context window and are available on Hugging Face.
Jais Speech transcribes Modern Standard Arabic and dialects from 19 Arab countries. It can identify dialects, add Arabic vowel marks and retain English or French words in their original scripts, supporting code-switching in spoken conversations.
According to the developers, the models were evaluated against 24 Arabic benchmarks that will be released publicly. The 7-billion-parameter model improved its average Arabic score from 56.4 on the K2 Horizon base model to 79.6 on Jais 3. The models include five training checkpoints and support deployment with vLLM and SGLang.
Source: Middle East AI News
