Core42, a subsidiary of G42 and provider of cloud and generative AI solutions, has announced the release of Jais 30B, the latest version of its open-source Arabic Large Language Model (LLM).
In an exclusive interview with Arabian Business, Andrew Jackson, EVP, Chief AI Officer, Core42 said, “I think the impact is amazing, potential is huge, but also the roadmap forward is massive.”
“We’ve demonstrated that we can do a technological breakthrough here, in this region from this region,” he added.
Jais 30B: Enhanced AI for Arabic Speakers
With over 30 billion parameters, Jais 30B builds upon the success of its predecessor, the 13 billion parameter Jais model, and aims to deliver an enhanced generative AI experience for Arabic speakers worldwide.
The model now provides 160 percent longer and more detailed answers in Arabic and a 233 percent increase in English, reflecting its enhanced language generation capabilities.
The development of Jais was a collaborative effort between Core42, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Cerebras Systems. The training process for Jais 13B was completed in a 21 days on the CG-1.
Trained on the Condor Galaxy-1 (CG-1), one of the world’s fastest AI supercomputers, the model boasts 4 exaFLOPS of training compute power, 54 million cores, and 64 nodes.
New AI Model Sets Benchmark
Jais 30B benefits from a significantly larger dataset, comprising 126 billion Arabic tokens, 251 billion English tokens, and 50 billion code tokens. This expanded training data has resulted in notable improvements across various performance indicators.
Additionally, Jais 30B exhibits improved performance in summarization (53 percent in Arabic and 85 percent in English) and formatting (130 percent in Arabic and 134 percent in English). These advancements position Jais 30B on par with monolingual English models and surpass most existing open-source models in Foundation Model evaluations.
To ensure the model’s reliability and effectiveness, Jais 30B underwent rigorous testing and validation using cross-model comparison, and human evaluations. The results demonstrated that the responses generated by Jais 30B’s outperformed those of Jais 13B in Arabic 96 percent of the time and in English 97 percent of the time.
In line with its commitment to responsible AI practices, the development team has further refined their processes and policies to mitigate biases and prevent the production of hateful or harmful content. The open-source nature of Jais facilitates transparency and enables greater scrutiny of the model’s outputs.
Responsible AI for Arabic
Jais’s advanced capabilities in the Arabic language domain have already impacted various sectors, including telecommunications, energy, education, healthcare, and marketing communications. The release of Jais 30B is set to drive further innovation in these fields, allowing organisations to leverage the power of generative AI for Arabic language applications.
When asked about the future of generative AI technology, Jackson said, “It’s a tough one to look into the future. But let me just try. I think Jais is in some way a flagbearer going into the future and is a really interesting demonstration of capability across language. So I think the language base of Jais will improve and increase, I think the quality of the model will improve.”
However, Jackson believes that, “there is no point competing with the likes of open AI.”
He states that “they have got such a head start in the industry and are backed by Microsoft,” which leads to Jackson’s belief of using a strategy of partnering with “the best” and finding their capability area accordingly.
Follow us on
