Jais 3 brings six open Arabic models, up to 375B

MBZUAI's Institute of Foundation Models (IFM) and Inception42 released Jais 3 on October 15, 2026. It is a family of six open Arabic language models, from 900 million to 375 billion parameters, plus a speech recognition model called Jais Speech. The release gives developers building for more than 400 million Arabic speakers free models that run anywhere from a phone to a server cluster.

MBZUAI is the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. Inception42 is the company that develops the Jais models with IFM. Cerebras Systems, a maker of AI chips, worked on the project too, Gulf News reports.

What is in the Jais 3 release

The six models share one training recipe and the same agent skills, Gulf News says. Agent skills mean the models can carry out tasks that use software tools, not just answer questions. Gulf News names five of the sizes and reports a context window of 512,000 tokens, which is how much text a model can take in at once.

Model	What the sources say
0.9B	The smallest model, sized to run on a phone
3.7B	Beats every open model up to 5B on Arabic and agent tasks, per its developers
7B	Highest average Arabic score among open models its size, per its developers
32B	Highest average Arabic score among open models its size, per its developers
375B	The flagship; leads open models on agent tasks, per its developers
Jais Speech	Speech recognition for Arabic and its dialects

Each model ships with five training checkpoints, saved snapshots from partway through training. RuntimeWire reports that the weights, checkpoints and evaluation code are published under Apache 2.0. That is a permissive license that allows commercial use. The models are on Hugging Face and support vLLM and SGLang, two popular tools for serving models.

Built on K2 Horizon, trained on real dialects

Jais 3 is built on K2 Horizon, IFM's fully open-source family of foundation models released in September 2026. Gulf News calls Jais 3 the first adaptation of K2 Horizon to a new language.

The training data covers Modern Standard Arabic, the formal written form, and spoken dialects. Those include Emirati, Gulf and North African Arabic. "We built Jais 3 for Arabic as it's really used," said Haonan Li, a senior research scientist at IFM, according to Gulf News.

Jais Speech transcribes Modern Standard Arabic and dialects spoken across 19 Arab countries. It handles speakers who switch between Arabic and English or Arabic and French mid-sentence. It can also name the dialect and add diacritics, the marks that show how Arabic words are pronounced.

How strong the claims are

Every performance figure comes from the developers. They tested on an audited suite of 24 Arabic benchmarks that they plan to release publicly. The 7B model's average Arabic score rose from 56.4 for the K2 Horizon base to 79.6 after Jais 3 training, Gulf News reports. The developers also say Jais Speech had the lowest average word and character error rates of the models compared, including commercial ones.

RuntimeWire notes that these claims rest largely on the developers' own evaluations. Independent tests have not been published yet. "Jais 3 makes Arabic AI more practical for business use," said Larry Murray, senior vice president of applied science at Inception42.

What this means for developers

• Start small. The 0.9B and 3.7B models fit on phones and modest servers. Try them for Arabic chat, search or form filling before you pay for the 375B model's hardware.
• Check the license yourself. RuntimeWire reports Apache 2.0, which allows commercial use. Confirm it on the Hugging Face model card before you ship a product.
• Test on your own dialect data. The benchmark claims are the developers' own. Run your users' real Gulf, Emirati or North African text through the model before you rely on it.
• Try Jais Speech for mixed-language audio. Call centers and voice apps often hear Arabic mixed with English or French. That is the case Jais Speech is built for.
• Use the checkpoints for fine-tuning. Five checkpoints per model let teams start from an earlier stage of training. That can help when you adapt a model to a narrow field such as law or health.