MBZUAI institute and Inception42 release six open Arabic Jais 3 models

The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42 have released Jais 3, a set of six open Arabic language models, together with a speech recognition model called Jais Speech. The release was announced in Abu Dhabi on 15 October 2026. The announcement describes Inception42 as G42’s sovereign agentic AI company.

The six language models range from 900 million to 375 billion parameters. According to IFM’s technical post, four are dense models, with 0.9 billion, 3.7 billion, 7 billion and 32 billion parameters, and two are mixture-of-experts models: one with 36 billion parameters in total and about 4 billion active, and one with 375 billion in total and about 23 billion active. All sizes except the smallest have a context window of 512,000 tokens; the smallest has 128,000. IFM says all model checkpoints and code are released under the Apache 2.0 licence. The models are on Hugging Face, with deployment support for vLLM and SGLang.

Built on the university’s own base models

Jais 3 is built on K2 Horizon, a family of six fully open-source foundation models that IFM released in September 2026, and is the first adaptation of K2 Horizon to a new language. Each Jais 3 model starts from the K2 Horizon model of the same size. IFM says continued pre-training used 201.45 billion tokens, of which 151.09 billion, or 75%, were Arabic. Its post says Jais 3 was developed in collaboration with Inception and Cerebras. Training compute and cost are not disclosed.

Jais Speech covers Modern Standard Arabic and dialects spoken in 19 Arab countries, handles switches into English or French, and can add Arabic vowel marks, according to the announcement. IFM is also publishing a suite of 24 audited Arabic benchmarks and the code to run them.

Claims and what is not disclosed

The developers say the 7 billion and 32 billion parameter models achieved the highest average Arabic scores among the open models they tested at those sizes, and that the 7 billion model’s average Arabic score rose from 56.4 for the K2 Horizon base to 79.6. They also say Jais Speech had the lowest average word and character error rates among the models compared, without publishing the rates in the announcement. These results come from the developers’ own evaluations, not from a peer-reviewed paper or an independent test.

Building an Arabic LLM is not a translation exercise.

Ashish Koshy, CEO, Inception42

The announcement says the Jais chat and mobile apps use the same models and that a Jais agentic platform will follow, without giving a date. It does not set out the commercial terms between MBZUAI and Inception42, pricing for the platform, or any named university, government or corporate customers. IFM was launched by MBZUAI in May 2025 and has labs in Abu Dhabi, Silicon Valley and Paris. MBZUAI, founded in 2019, has nearly 1,000 students from more than 70 countries, according to the release.

What it means

TED analysis:

• Research to product: MBZUAI is building base models in house, adapting them for a language and passing them to a partner company to deploy. Other research universities in the Gulf will be compared with this route from research to use.
• Campus use: An Apache 2.0 licence lets Arabic-medium universities in the UAE, Saudi Arabia, Qatar and elsewhere run Jais 3 on their own infrastructure for teaching, administration and research, subject to their own data rules. The smaller models matter most here because they need far less hardware.
• Evidence: The benchmark gains are the developers’ own figures. Universities and ministries considering adoption will want independent evaluation, and the published benchmark suite lets others check the claims.
• Suppliers: Edtech and student services companies selling to Arabic-speaking universities now have an open, locally developed model to build on, which weakens the case for paying for closed models in Arabic-language products.