MBZUAI and Inception42 launch six open Jais 3 models and Jais Speech

MBZUAI and Inception42 have launched Jais 3 and Jais Speech. The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42 released the models on 15 October 2026. Jais 3 is a set of six open language models built for Arabic. The models range from 900 million to 375 billion parameters. Jais Speech is a speech recognition model for Arabic and its dialects. Both are available on Hugging Face. The models are trained on Modern Standard Arabic and regional dialects spoken in 19 Arab countries. They handle switches between Arabic and English or French in the same sentence. Jais Speech can identify the dialect and add vowel marks to the text. The developers say the combination can turn spoken instructions into practical work such as transcribing a conversation or drafting a document or presentation. Jais 3 is built on K2 Horizon, the open foundation model fleet that the Institute of Foundation Models released in September 2026. The six models include dense models at 0.9 billion, 3.7 billion, 7 billion and 32 billion parameters, plus two mixture-of-experts models. The largest is a 375-billion-parameter model with 23 billion active parameters. The partners developed the models with Cerebras Systems. The models and checkpoints use the Apache 2.0 license. The developers evaluated the models on an audited suite of 24 Arabic benchmarks that they are releasing publicly. They say the largest model leads open models on agentic tasks and matches the strongest on Arabic understanding. Jais Speech recorded the lowest average word and character error rates among the speech systems they compared. Those results are developer-reported. No independent replication is included in the announcement. The Jais family first launched in Abu Dhabi in August 2023. Jais 3 moves from generating text to carrying out multi-step tasks. The partners have not listed the number of developers who have downloaded the models or the number of production systems already using them. They have not published the cost of training or the number of researchers who worked on the release. For a developer the practical change is that the model files are available now. A team can download a model that runs on a phone or a model that needs a larger server. The partners have not listed any government or company that has already put Jais 3 into daily use. The value depends on whether later reports show the models are used for real work in Arabic. Other teams have also released Arabic language models. Some are fine-tuned versions of models built outside the region. Jais 3 is trained from the start on Arabic dialects and is released as open weights. The comparison is between models that treat Arabic as an added language and a model built for Arabic first. The announcement does not list any new funding raised or any change in the number of researchers at the institute. It lists the model sizes and the benchmark results the developers measured. A reader can see the intended scope. A reader cannot yet see how many real users the models have or how they perform on tasks the developers did not test. The partners have not named the government agencies or companies that will use the models first. Those names would show who is already testing the models on real Arabic speech and documents. Without them the release remains a set of files on Hugging Face. The models support a context length of 512,000 tokens. That length lets a single request include a long document or a long transcript. The partners have not listed how many real tasks have already used that full length. If the partners publish the number of downloads and the accuracy of Jais Speech on speech collected in the UAE, the release has a measurable follow-through. If the next update is only a further model card, the change for developers remains the same as the initial announcement. The figures missing are the number of production users and the measured error rate on local speech.