The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42 on 15 October 2026 introduced Jais 3, a fleet of models built for Arabic. The release, set out in an IFM press statement, pairs six open language models, from 900 million to 375 billion parameters, with Jais Speech, a speech-recognition model for Arabic and its dialects.
All of the models are built on K2 Horizon, IFM’s latest fleet of fully open-source foundation models, launched in September 2026. The statement describes Jais 3 as the first adaptation of K2 Horizon to a new language. The work was done in collaboration with Cerebras Systems.
According to the statement, the models are trained to understand Modern Standard Arabic and regional dialects “from the Gulf to North Africa,” along with poetry, cuisine and Islamic knowledge. That understanding, it says, matters for “the more than 400 million people who speak Arabic,” and for hospitals, companies and government offices across the Middle East and North Africa.
IFM and Inception42 tested the models on their own audited suite of 24 Arabic benchmarks. The statement says the largest Jais 3 model matches the best open models on Arabic understanding and leads them on agentic tasks, and that Jais Speech leads every model compared on Arabic speech recognition. It also says the 3.7-billion-parameter model outperformed every competing open model of up to 5 billion parameters on Arabic and agentic tasks, and that the 7-billion- and 32-billion-parameter models earned the highest average Arabic scores among open models tested at their sizes. At 7 billion parameters, the average score on the institutes’ Arabic evaluations rises from 56.4 for the K2 Horizon base model to 79.6 for Jais 3.
The press statement says all six language models share a context window of 512,000 tokens, “the longest of any Arabic-native model.” IFM’s technical blog notes an exception: the 0.9-billion-parameter model has a 128,000-token window.
Jais Speech, the statement says, transcribes Modern Standard Arabic and dialects spoken across 19 Arab countries in noisy, real-world conditions. It follows speakers as they switch between Arabic and languages such as English or French, keeping those words in their original script, identifies the dialect being spoken, and can add Arabic vowel marks. In an evaluation across six public test sets, the institutes say Jais Speech recorded the lowest average word and character error rates of the models compared, including commercial speech-recognition services.
“Arabic AI needs to understand the language in its everyday context, which varies by geography and across generations,” Haonan Li, a senior research scientist at IFM, said in the statement. “For example, we need to account for the many nuances in spoken Arabic, including when speakers slip into English or French mid-sentence and back again. We built Jais 3 for Arabic as it’s really used.”
“Jais 3 makes Arabic AI more practical for business use,” said Dr. Larry Murray, SVP of Applied Science at Inception42. “Stronger reasoning and instruction following, combined with tool use, provide a foundation for multi-step agentic workflows. With open models across a range of sizes, organisations can match these capabilities to their computing budgets and deployment needs.”
“Jais 3 brings deep Arabic understanding to speech, writing and agents,” said Richard Morton, vice president and executive director of IFM. “It is built for the way Arabic is spoken, the way it is written, and the work people need done in it.”
Ashish Koshy, CEO of Inception42, said: “Jais has defined what an Arabic large language model can be since the first release, and Jais 3 is the clearest step up yet… That is the work we and IFM have done together across every generation of Jais, and it is why, by bringing together 6 new models, Jais 3 now leads on Arabic across every model size we have built.”
The models and Jais Speech are available on Hugging Face, with deployment support for vLLM and SGLang, the statement says. The technical blog adds that the models and code are released under Apache 2.0, together with training-stage checkpoints, audited Arabic benchmarks and evaluation code.