The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42, G42's sovereign agentic AI company, have introduced Jais 3, a fleet of six open language models built for Arabic and paired with Jais Speech, a speech recognition model covering Arabic and its dialects. The models range from 900 million to 375 billion parameters and are built on K2 Horizon, the fully open fleet that IFM released in September, making Jais 3 the first adaptation of that foundation to a new language.

The release is aimed squarely at deployment. Jais 3 is trained on Modern Standard Arabic and the dialects spoken from the Gulf to North Africa, carries a context window of 512,000 tokens, the longest of any Arabic-native model, and can plan multi-step tasks, search the web and produce Arabic documents, spreadsheets and presentations when connected to software tools. On an audited suite of 24 Arabic benchmarks, the largest model matched the best open models on Arabic understanding and led them on agentic tasks, while Jais Speech, which transcribes speech from 19 Arab countries and follows speakers as they switch into English or French, recorded the lowest average error rates of the models compared, including commercial speech recognition services.

Ashish Koshy, chief executive of Inception42, said building an Arabic large language model was not a translation exercise but meant training on the dialects people actually speak and the culture the language carries. The models are available on Hugging Face with deployment support for vLLM and SGLang, and each ships with five training checkpoints and the audited Arabic benchmarks, allowing enterprises and governments to adapt the models and decide where to deploy them in line with sovereign AI requirements.