UAE News: Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42 company Inception42 released the Jais 3 Arabic AI models on 15 October 2026, a family of six open systems from 900 million to 375 billion parameters, plus a speech model built to understand dialects from 19 Arab countries.
Key facts
- Jais 3 was developed by MBZUAI’s Institute of Foundation Models (IFM) and Inception42 with US chipmaker Cerebras Systems.
- All six models share a 512,000-token context window; the smallest, at 900 million parameters, is designed to run on a phone.
- Jais Speech transcribes Modern Standard Arabic and regional dialects, and keeps English or French words in their original script when speakers switch mid-sentence.
- The models are free to download on Hugging Face and are built on K2 Horizon, IFM’s open-source base fleet released in September 2026.
What is in the Jais 3 Arabic AI release?
Jais 3 is a fleet of models. The developers say every size uses the same training recipe and the same agent functions, so a company can prototype on a small model and move up without rewriting its tools. Each model ships with five training checkpoints, which lets outside researchers study how the system changed during training.
| Component | Size or feature | Intended use (developer description) |
|---|---|---|
| Smallest model | 900 million parameters | On-device use, including phones |
| Small model | 3.7 billion parameters | Low-cost Arabic and agent tasks |
| Mid-size models | 7 and 32 billion parameters | Enterprise assistants, instruction following |
| Largest model | 375 billion parameters | Agentic tasks, complex Arabic understanding |
| Jais Speech | Dialects of 19 Arab countries | Transcription, dialect identification, voice commands |
The speech model can also identify which dialect is being spoken and add vowel marks to transcripts. That matters for Arabic, where written text usually omits short vowels. Paired with the language models, it lets Jais 3 act on spoken instructions in Arabic or English.
How strong are the benchmark claims?
The headline numbers come from the developers. According to the 15 October statement carried by Zawya and Gulf News, the 7-billion-parameter model’s average Arabic score rose from 56.4 for the K2 Horizon base to 79.6 after Arabic training. The 3.7-billion model is said to beat every open model up to 5 billion parameters on Arabic and agentic tasks, and Jais Speech reportedly posted the lowest average word and character error rates of the systems compared, including commercial services.
Those results rely on an audited suite of 24 Arabic benchmarks that IFM says it will publish. Independent replication had not been reported by 17 October. Technical coverage by AlphaSignal, drawn from IFM’s documentation, also noted that about 59% of the Arabic training tokens are synthetic or machine-translated, and that the largest model trails some closed commercial systems on Arabic long-context reasoning.
“We built Jais 3 for Arabic as it’s really used,” said Haonan Li, senior research scientist at IFM. Ashish Koshy, chief executive of Inception42, put the commercial case directly: “Building an Arabic LLM is not a translation exercise.”
Sovereign AI and the Arabic-speaking market
The first Jais model, launched in Abu Dhabi in 2023, had 13 billion parameters. Three years on, the top model is almost 30 times larger. The developers pitch Jais 3 at governments and companies that want to keep data and deployment under their own control, citing more than 400 million Arabic speakers as the potential user base.
Open weights are the point. A ministry in Cairo or a bank in Riyadh can download the models, adapt them and host them on local servers instead of sending queries abroad. Richard Morton, executive director of IFM, said Jais “has shown what the UAE’s long-term investment in open AI research can deliver for the Arabic-speaking world.”
The release fits a wider push by BRICS members to build their own AI stacks rather than rent them. China’s spending surge, covered in our report on China’s R&D and the Nature Index, is the largest example; the UAE is betting on openness and language specialisation at a smaller scale.
Our view: the Jais 3 Arabic AI family is a serious contribution to Arabic computing because it is open, documented and offered in sizes that run on ordinary hardware. Its performance claims deserve the same scrutiny as any vendor benchmark until outside groups test the 24-task suite once it is published.
Follow more stories on our Technology and Innovation pages.
FAQ
Is Jais 3 free to use?
The models and Jais Speech are published on Hugging Face as open models. Users should check each model’s licence page before commercial deployment.
Which Arabic dialects does Jais 3 cover?
The developers say training covers Modern Standard Arabic and regional dialects including Emirati, Gulf and North African Arabic, while Jais Speech handles dialects from 19 Arab countries.
Who built it?
MBZUAI’s Institute of Foundation Models and Inception42, part of Abu Dhabi’s G42 group, with Cerebras Systems.
Sources: Zawya (IFM and Inception42 statement); Gulf News; Middle East AI News; AlphaSignal.
