Most AI models speak Arabic as a second language. They were built in English first, and Arabic was added later — which is why they stumble on dialects, on Egyptian or Moroccan speech, and on the everyday habit of switching between Arabic and English mid-sentence.
Jais 3, launched on 15 October 2026 by Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42, reverses that order. It is Arabic first, and it is built to do work, not just hold a conversation.
A family, not a single model
Jais 3 is six open language models sharing one training recipe, from a 0.9-billion-parameter model designed for phones up to a 375-billion-parameter flagship for data centres. The two largest use a mixture-of-experts design — the flagship stores 375 billion parameters but activates only 23 billion per token, keeping running costs sane. Five of the six have a 512,000-token context window (the smallest has 128K), which IFM says is the longest of any Arabic-native model.
All six are on Hugging Face under the Apache 2.0 licence, so developers can use them commercially and self-host, from a phone app to a GPU cluster. They are built on K2 Horizon, IFM's open foundation model fleet from September 2026, adapted with an Arabic-focused training recipe covering Modern Standard Arabic and regional dialects, plus the cultural knowledge around the language.
The speech part is the practical win
Alongside the text models comes Jais Speech (Jais-ASR), a speech recognition system for Arabic and its dialects across 19 Arab countries. It transcribes dialects, accepts spoken instructions in Arabic or English, handles code-switching between Arabic and English or French, and applies Arabic diacritics (vowel marks) directly to live transcripts — something most speech tools skip, and which makes transcripts far more useful for documents and subtitles.
Built to act
IFM's pitch is that Jais 3 works as an agent: it plans multi-step tasks, calls tools, edits Arabic documents, searches the web and hands back a finished result. Launch examples include reading an Arabic contract, updating a spreadsheet and producing presentations. It is also designed to follow instructions closely — format, length and constraints as set by the user.
The honest caveat
Every headline benchmark number comes from IFM's own evaluations across 24 public Arabic benchmarks and has not yet been independently replicated. The most eye-catching: at 7 billion parameters, the average Arabic score rose from 56.4 for the K2 Horizon base to 79.6 after Arabic adaptation. That is a developer-reported claim, not a settled fact — and the right way to judge Jais 3 is to test it on your own dialect and documents.
Who should care
If your documents, customers and meetings are in Arabic, Jais 3 is a practical upgrade over an English-first chatbot — especially for transcription and document work. Developers building for the Arab world get six self-hostable sizes with no API bill. If you work only in English, Jais 3 still matters as a signal: language-native open models, built by countries for their own languages, are arriving fast.
This is a condensed explainer. The full breakdown — the complete model table, the benchmark picture, pros and cons, and FAQ — is on NeurAITool: Jais 3: The UAE's Open Arabic AI Family Explained.
Top comments (0)