UAE launches Jais 3, Arabic-native AI for speech, writing and agents

The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42 have unveiled Jais 3, a collection of Arabic-native AI models designed for speech, writing, and agents. Jais 3 represents the latest iteration of the Jais family, which was initially introduced in Abu Dhabi.

This advanced AI system is capable of understanding and generating Arabic text across various dialects, as well as performing practical tasks such as transcribing conversations and creating documents and presentations. The Jais 3 family is comprised of six open language models, ranging from 900 million to 375 billion parameters, paired with Jais Speech, a speech recognition model tailored for Arabic and its diverse dialects.

These models are built on K2 Horizon, a fully open-source foundation model fleet released by IFM in September 2026. According to Haonan Li, a senior research scientist at IFM, Arabic AI must grasp the nuances of the language as it is used in everyday contexts, which vary significantly across different regions and generations. He emphasized that Jais 3 is designed to mirror the way Arabic is truly spoken and written.

Jais 3 was evaluated against a rigorous suite of 24 Arabic benchmarks, and the results demonstrated its superiority in agentic tasks and Arabic understanding compared to other open models. The largest Jais 3 model outperformed competing models in terms of agentic tasks, while the smallest model led open models in performance. Jais Speech, an integral component of Jais 3, boasts the lowest error rates among all models, excelling in transcribing Arabic speech in real-world conditions.

This includes maintaining the original script and identifying the dialect being spoken. Furthermore, Jais 3 demonstrates strong performance in following complex Arabic instructions, making it a practical solution for business applications. The six models in the Jais 3 family cater to different computational requirements, with options ranging from a 900-million-parameter model to a 375-billion-parameter flagship.

All models share the same training recipe and agent capabilities, including a context window of 512,000 tokens, the longest of any Arabic-native model. Jais 3 is built upon the foundation of K2 Horizon, an open-source model fleet, which was adapted for the first time to a new language. Developed by IFM at MBZUAI, Inception42, and with the collaboration of Cerebras Systems, Jais 3 and Jais Speech are now available on Hugging Face, with deployment support for vLLM and SGLang.