Jais 3: The UAE's Open Arabic AI Family Explained

Most of the world's best AI models speak Arabic as a second language. They were built in English first, then taught Arabic afterwards — and it shows, especially when someone switches dialect mid-sentence or slips an English word into an Arabic conversation. Jais 3, launched on 15 October 2026 by Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42, flips that order: it is Arabic first, built for the way more than 400 million people actually speak, write and work in the language.

Jais 3 is not one model but a family of six open language models, from a 900-million-parameter model for phones to a 375-billion-parameter flagship for data centres, plus an Arabic speech system called Jais Speech. All are on Hugging Face under the Apache 2.0 licence, so developers and businesses can use them commercially without a per-token API fee.

What Is Jais 3?

Jais 3 is the third generation of the Jais family, the UAE's flagship Arabic AI project. It was developed by MBZUAI's Institute of Foundation Models (IFM) together with Inception42, in collaboration with Cerebras Systems. Every model in the family is built on K2 Horizon, IFM's fully open-source foundation model fleet released in September 2026, then adapted with an Arabic-focused training recipe.

That adaptation is the whole point. Jais 3 is trained on Modern Standard Arabic and regional dialects — from Emirati and Gulf Arabic to Egyptian and North African varieties — plus the cultural knowledge wrapped up in the language, including its poetry, cuisine and Islamic knowledge. As IFM senior research scientist Haonan Li put it at launch: "We built Jais 3 for Arabic as it's really used."

The Six Models, From Phone to Data Centre

The family shares one training recipe across six sizes, so a developer can pick the model that fits the hardware and budget available. Here is the line-up:

Model	Parameters	Context window	Best for
Jais-3-0.9B	0.9 billion (dense)	128K tokens	Phones and edge devices
Jais-3-3.7B	3.7 billion (dense)	512K tokens	Laptops and small servers
Jais-3-7B	7 billion (dense)	512K tokens	Everyday business and developer use
Jais-3-32B	32 billion (dense)	512K tokens	Serious production workloads
Jais-3-MoVA-36B-A4B	36 billion total, 4 billion active per token	512K tokens	Efficient high-performance serving
Jais-3-375B-A23B	375 billion total, 23 billion active per token	512K tokens	Frontier-class work on GPU clusters

The two largest models use a mixture-of-experts design: they store a huge number of parameters but only activate a fraction for each token. The flagship, for example, activates just 23 billion of its 375 billion parameters per token. The 512,000-token context window on five of the six models — the 0.9B model has 128K — is, according to IFM, the longest of any Arabic-native model.

One practical note: Jais 3 is open-weight under Apache 2.0, which is genuinely permissive. It sits in the same broad open-model wave as Meta's 816B Muse Spark Open and the models compared in our DeepSeek vs Qwen vs Kimi open-model comparison — but where those are general-purpose models, Jais 3 is a specialist, and specialists usually win on their home turf.

Jais Speech: Arabic as People Actually Speak It

The most immediately useful part of Jais 3 may not be a text model at all. Jais Speech (also called Jais-ASR) is a speech recognition system built for Arabic and its dialects across 19 Arab countries. It transcribes Modern Standard Arabic and regional dialects, accepts spoken instructions in Arabic or English, and handles code-switching — switching between Arabic and English or French mid-sentence and back again.

It also applies Arabic diacritics (vowel marks) directly to live transcripts, which most speech tools skip. Written Arabic normally omits vowel marks, so a transcript without them can be ambiguous; adding them makes transcripts far more useful for documents, subtitles and records. IFM reports that Jais Speech recorded the lowest average word and character error rates across 19 Arab countries in its own evaluations — a developer-reported claim, since independent replication has not yet appeared. For a broader look at voice AI, see our tested roundup of the best AI voice generators.

What Jais 3 Can Actually Do

IFM's pitch is that Jais 3 is built to act, not just chat. The models are trained as agents: they plan multi-step tasks, call software tools, edit Arabic documents, search the web and hand back a finished result, all while talking to the user in Arabic. Launch examples include reading an Arabic contract, updating a spreadsheet, transcribing a conversation, and producing documents and presentations.

The team also highlights instruction-following: Jais 3 is designed to stick to the format, length and constraints a user sets. For anyone writing in Arabic professionally, that combination — a model that follows a brief and finishes a document — is more valuable than raw chat fluency. Our guide to the best AI writing tools covers the English-first alternatives for comparison.

How Good Is It? The Honest Benchmark Picture

Here is where a careful reader should slow down. The headline numbers all come from IFM's own evaluations across 24 public Arabic benchmarks, and they have not yet been independently replicated. That does not make them wrong — the benchmarks are public, and independent testing will follow — but it does mean treating them as the developers' claims rather than settled facts.

With that caveat, the claims are striking. At 7 billion parameters, Jais 3's average Arabic evaluation score rose from 56.4 for its K2 Horizon base model to 79.6 after the Arabic adaptation. IFM says the 375B flagship leads open models on agentic tasks and matches the strongest models on Arabic understanding, and that the 7B model beats some open models more than four times its size on agentic benchmarks. Independent reviewers who have examined the model cards note the picture is more mixed than the headlines suggest, which is normal for a launch week. The sensible verdict: promising, genuinely open, and worth testing on your own Arabic workloads rather than taking any single score on trust.

Who Is Jais 3 For?

Arabic-speaking businesses and government teams. If your documents, customers and meetings are in Arabic, an Arabic-native model with dialect-aware speech is a practical upgrade over an English-first chatbot, especially for transcription, document work and customer service.

Developers building for the Arab world. Six sizes under Apache 2.0, on Hugging Face with training checkpoints, recipes and evaluation materials, means you can self-host, fine-tune and ship without an API bill or vendor lock-in — from a phone app using the 0.9B model to a server deployment of the 32B.

Researchers and open-AI watchers. Jais 3 tests a bigger question: can a national programme produce a language-native open model that beats the global giants on their own turf? The UAE's National AI Strategy 2031 is the backdrop, and Jais 3 is its most visible open-AI output so far.

Probably not you, yet, if you work only in English. Jais 3 is multilingual, but its edge is Arabic. English-only users will get more from the general-purpose models we cover elsewhere on NeurAITool.

Jais 3 vs the English-First Giants

The honest comparison is not "Jais 3 versus ChatGPT" in general — it is "Jais 3 versus ChatGPT in Arabic." On English benchmarks, the frontier closed models still lead. On Arabic understanding, dialect handling, Arabic speech transcription and Arabic agent work, Jais 3's design gives it a real shot at leading the open field. The gap that matters most in practice is dialect and speech: most leading models handle formal Modern Standard Arabic passably but stumble on Egyptian, Moroccan or Emirati speech. That is exactly the gap Jais 3 was built to close, and it is the thing to test first if you try it.

Pros and Cons

Pros: genuinely Arabic-native, with dialects and code-switching; six sizes from phone to data centre; Apache 2.0 licence with commercial use; 512K context on most models; dedicated Arabic speech with diacritics; agent skills built in; weights, checkpoints and recipes published.

Cons: benchmark claims are developer-reported and not yet independently verified; the flagship needs serious GPU hardware (IFM's reference deployment is 16 H100 GPUs, or a single eight-GPU H200 node); English performance is not the point; the ecosystem of Arabic fine-tunes and tools around Jais 3 is still young.

Verdict: Should You Try Jais 3?

Yes — if you work in Arabic. Jais 3 is the most serious open Arabic AI family yet released, and its combination of dialect-aware speech, long context and agent skills solves real problems that English-first models handle poorly. Start with the 7B model for general work or Jais Speech for transcription, and judge it on your own dialect and documents — not on launch-week scores. If you work only in English, watch it instead: Jais 3 points to a future of language-native open models built for their own languages, rather than one English-first model for everyone.

Sources: the official IFM press release and technical blog post (IFM press release, IFM technical blog), plus WAM/Gulf coverage of the 15 October 2026 launch. Benchmark figures are IFM's own evaluations, flagged as such above.

Frequently Asked Questions

What is Jais 3?

Jais 3 is a family of six open Arabic-language AI models from MBZUAI's Institute of Foundation Models and Inception42, launched on 15 October 2026, plus Jais Speech for Arabic speech recognition. The models range from 0.9 billion to 375 billion parameters.

Is Jais 3 free to use?

The models are released under the Apache 2.0 licence on Hugging Face, which permits commercial use. You still need hardware to run them — the small models run on modest devices, while the 375B flagship needs a GPU cluster.

How is Jais 3 different from ChatGPT in Arabic?

ChatGPT and most leading models are English-first with Arabic added later. Jais 3 is Arabic-native: trained on Modern Standard Arabic and regional dialects, with speech that handles dialect variation and code-switching between Arabic, English and French.

Can Jais 3 run on a phone?

The smallest model, Jais-3-0.9B, is designed for phones and edge devices, while larger models need a laptop, server or data centre.