What happens when you dictate a request to an AI assistant in your own dialect, slip into English for a technical term mid-sentence, and slip back out? For most models, the answer is confusion. For Jais 3, it is the exact problem the family was built to solve. On Thursday, October 15, 2026, the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), together with Inception42 and in collaboration with Cerebras Systems, launched the full Jais 3 lineup: six open-weight, Arabic-native language models plus Jais Speech, a speech-recognition model covering Modern Standard Arabic and the dialects of 19 Arab countries. This guide, updated October 15, 2026, explains what launched, why it differs from translated versions of global chatbots, and how to start using it today — free of charge.
What Is Jais 3, and Why Does "Arabic-Native" Matter?
Most frontier models treat Arabic as a second-class citizen: the input is internally mapped toward English, reasoned about, and the output is translated back. The result is text that is grammatically defensible but stiff, overly formal, and oddly disconnected from how people actually speak. Jais 3 flips that architecture. The family is built on IFM's K2 Horizon foundation fleet, released in September 2026, but trained natively on Modern Standard Arabic and regional dialects — including the way speakers actually mix languages in a single sentence.
"Arabic AI needs to understand the language in its everyday context, which varies by geography and across generations," said Haonan Li, senior research scientist at IFM, in the launch announcement. "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."
The Six Models: Sizes and Best-Fit Uses
All six models are available now under the Apache 2.0 license — which permits commercial use — on Inception42's official Hugging Face page:
| Model | Size | Best fit |
|---|---|---|
| Jais-3-0.9B | 0.9B parameters | Low-resource devices and quick experiments |
| Jais-3-3.7B | 3.7B parameters | Embedded apps and phones |
| Jais-3-7B | 7B parameters | Small projects and lightweight agents |
| Jais-3-32B | 32B parameters | Business and serious production use |
| Jais-3-MoVA-36B-A4B | 36B total, 4B active | High speed at low operating cost |
| Jais-3-375B-A23B | 375B total, 23B active | Flagship: top accuracy, 512K-token context |
The flagship Jais-3-375B-A23B uses a Mixture-of-Experts design: it stores 375 billion parameters but activates only 23 billion per token, pairing a vast knowledge base with reasonable inference speed. Its 512K-token context window means it can absorb very long Arabic documents — entire reports, transcripts, or reference collections — in a single session.
Source: Official model card on Hugging Face
Jais Speech: The Ear That Understands 19 Countries
The second half of the launch is Jais Speech, and for anyone building voice products it is the more consequential piece. Its declared capabilities:
- Understands Modern Standard Arabic plus dialects from 19 Arab countries, with emphasis on Emirati, Gulf, and North African speech.
- Automatically identifies which dialect is being spoken.
- Preserves English and French words inside an Arabic sentence in their original script instead of mangling them — a chronic failure point for conventional transcription tools.
- Adds vowel marks (tashkeel) to transcripts automatically, useful for educational and legal documentation.
The practical outcome, according to the developers: you speak an instruction in Arabic or English, and the system converts it into action — transcribing conversations, drafting documents and presentations, and executing multi-step tasks.
Source: Official model card on Hugging Face
How to Start Using Jais 3 Today
- Open the official Inception42 organization page on Hugging Face and pick the model that matches your workload from the table above.
- Test it immediately through the free inference widget on the model card — no download required.
- Developers: the models are compatible with the transformers library, and a compressed FP8 build of the flagship, Jais-3-375B-A23B-FP8, is available for cheaper serving.
- For voice use cases, follow IFM's documentation for integrating Jais Speech with the language models.
Size your hardware before you commit. The 0.9B model runs comfortably on an ordinary laptop, making it ideal for prototypes. The 7B and 32B models need a single strong GPU or a small cloud instance. The flagship's sparse architecture means its effective running cost resembles a model with roughly 23 billion parameters, despite its far larger stored knowledge — the economic trick that makes the top of the lineup commercially viable rather than a tech demo.
This is also the first release since the family debuted in August 2023 — followed by Jais 2 in December 2025 — to ship this many open sizes alongside a speech component, putting it in direct competition with major open-weight releases such as GLM-5.3 open weights and Tencent's Hy4 with 770B parameters.
What This Means for You
- Students and researchers get a model that handles dialectal input and long Arabic references without an English intermediary distorting meaning — the huge context window is tailor-made for summarizing lengthy source material.
- Developers and startups can build commercial products on top of Apache 2.0-licensed weights with zero licensing fees, and the Mixture-of-Experts design keeps serving costs in check.
- Education and media organizations can transcribe lectures and meetings in regional speech more accurately than general-purpose tools — a trend visible across the Arabic voice landscape, from Gemma 4 for real-time voice AI to Gemini 3.8 Live supporting 97 languages.
- Privacy-sensitive users can self-host: open weights mean your data never has to leave your own infrastructure.
Quick Comparison: Jais 3 vs the Big Chatbots
| Criterion | Jais 3 | ChatGPT | Gemini |
|---|---|---|---|
| Native Arabic (dialects, code-switching) | Yes — the foundation | Translated pipeline | Partially translated |
| Open weights | Yes — Apache 2.0 | No | No |
| Private self-hosting | Yes | No | No |
| Arabic speech component | Jais Speech, 19 countries | Yes (not Arabic-specialized) | Yes (not Arabic-specialized) |
| Best for | Privacy and Arabic precision | Broadest tool ecosystem | Google integration |
Honest Limitations
The benchmarks published at launch come from the development team itself and await independent verification over the coming weeks. Open models require technical skill to self-host, and hosted inference providers for this family are still limited compared with the giants. Jais Speech is a speech-recognition-plus-language system, not a full replacement for cloud assistants that combine models with live web search and tool integrations. Dialect coverage also varies: Gulf and Emirati speech sits closest to the training distribution, and more distant dialects may produce weaker results.
Frequently Asked Questions
Is Jais 3 really free?
The six models' weights are published under Apache 2.0, which allows commercial use. You only pay if you rent servers or a hosted inference provider to run them.
How is it different from using ChatGPT in Arabic?
ChatGPT routes your request through an internal English-centric process, while Jais 3 is trained directly on Arabic and its dialects, preserving local meaning and mid-sentence language switching.
Does it support my dialect?
Coverage spans 19 Arab countries with emphasis on Emirati, Gulf, and North African dialects. Accuracy varies by dialect, so test with your own audio samples.
Where can I try the flagship model directly?
On its official Hugging Face model card, with a lighter FP8 build available as Jais-3-375B-A23B-FP8.
Can I build a commercial product on it?
Yes — the license explicitly permits it, and the sparse architecture was chosen precisely to lower commercial operating costs.