Mohamed bin Zayed University of Artificial Intelligence’s Institute of Foundation Models (IFM) and Inception42 have released Jais 3, a fleet of Arabic-native models, together with Jais Speech, an Arabic speech-recognition model. A press release carried on Zawya is dated 15 October 2026; IFM’s product page is at ifm.ai/jais.
The package pairs six open language models, from 900 million to 375 billion parameters, with Jais Speech. All are built on K2 Horizon, IFM’s fully open-source foundation model fleet released in September 2026. The models are available today on Hugging Face, with deployment support for vLLM and SGLang. Jais 3 was built in collaboration with Cerebras Systems.
The press release does not state the licence terms for the weights. Readers should check Hugging Face model cards before assuming Apache, MIT or another OSI licence.
What Jais 3 is meant to do
IFM and Inception42 say Jais 3 is built for Arabic from the start — not English-first models with Arabic bolted on — for more than 400 million speakers. Training covers Modern Standard Arabic and regional dialects from the Emirati and Gulf dialects to North African Arabic, plus cultural knowledge the language carries.
Jais Speech is described as covering dialects spoken across 19 Arab countries, handling code-switching into English or French (keeping those words in their original script), identifying dialect, and optionally adding Arabic vowel marks. Paired with the language models, it is meant to take spoken instructions in Arabic or English and act on them.
All six language models share the same training recipe and agent capabilities, with a context window of 512,000 tokens — which the release calls the longest of any Arabic-native model. That length claim is theirs.
Benchmark claims — their evaluations
IFM and Inception42 evaluated the models on an audited suite of 24 Arabic benchmarks, which they say they are releasing publicly. Figures to treat as their own evaluations, not independent third-party audits:
- At 7 billion parameters, the average Arabic score rises from 56.4 for the K2 Horizon base model to 79.6 for Jais 3
- The largest Jais 3 model is said to lead open models on agentic tasks and match the strongest on Arabic understanding
- Jais Speech is said to have recorded the lowest average word and character error rates of any model compared, including commercial ASR services
- The 3.7B model is said to have outperformed every competing open model up to 5B on Arabic and agentic tasks; 7B and 32B are said to lead open models at their sizes on average Arabic scores
IFM’s site publishes detailed comparison tables against other open and closed models; those tables are vendor-run leaderboards.
Checkpoints and deployment
Each model ships with five training checkpoints, so researchers can trace development and build from intermediate stages. The release says government entities and enterprises in the UAE and beyond can adapt the models and decide where to deploy them, in line with sovereign AI requirements. Sizes range from a 900-million-parameter model described as small enough to run on a phone to the 375-billion-parameter flagship.
What the people said
Ashish Koshy, CEO of Inception42, said: “Building an Arabic LLM is not a translation exercise. It means training on the dialects people actually speak, the way they write, and the culture the language carries.”
Haonan Li, senior research scientist at IFM, stressed everyday context, geography and generations, including speakers who slip into English or French mid-sentence. Richard Morton, IFM’s vice president and executive director, tied the release to the UAE National AI Strategy 2031. Dr Larry Murray, SVP Applied Science at Inception42, said stronger reasoning, instruction following and tool use provide a foundation for multi-step agentic workflows across compute budgets.
Why it matters
Open Arabic-native stacks remain thinner than English-centric open models. Jais 3’s pitch is a full range of sizes on a shared K2 Horizon base, plus speech and agent tooling, released for anyone to download. That sits alongside other open-weight drops TSN has covered, such as StepFun’s Step 5 Preview — which shipped under its own licence, not Apache 2.0. For Jais 3, licence terms are simply not stated in the announcement materials checked here.
Caveats
Benchmark leadership, ASR rankings and the “longest context” claim are IFM/Inception42’s. Licence is not stated in the press release. Availability on Hugging Face does not by itself define commercial terms or redistribution rights.
What to watch next
- Licence text on the Hugging Face model cards
- Third-party reproduction of the 24-benchmark suite
- Whether enterprises publish on-prem Jais 3 deployments in the UAE and elsewhere
Sources
- Zawya press release, “MBZUAI’s Institute of Foundation Models and Inception42 launch Jais 3: Arabic-native AI for speech, writing and agents”, 15 October 2026. https://www.zawya.com/en/press-release/companies-news/mbzuais-institute-of-foundation-models-and-inception42-launch-jais-3-arabic-native-ai-for-speech-writing-and-agents-1672443
- IFM, Jais 3 product page. https://ifm.ai/jais/
