Two new Arabic models were released in ten days by two Abu Dhabi institutions. That is good news for any organisation that wants Arabic AI running on its own premises, provided it knows what to measure before it chooses.

What was released

On 6 October the Technology Innovation Institute (TII) announced Falcon-Emirati, a 7-billion-parameter language model built on Falcon-H1-Arabic and purpose-built for Emirati Arabic, together with a 1.6-billion-parameter speech recognition model (Falcon-ASR) and an Arabic text extraction model (Falcon-OCR-Arabic). TII reports 84.83% on Alyah, a native Emirati Arabic benchmark of its own making.

On 15 October the Institute of Foundation Models at MBZUAI and Inception42 released the Jais 3 family: six open language models from 900 million to 375 billion parameters, built on K2 Horizon, plus Jais Speech, an Arabic speech recognition model covering dialects spoken across 19 Arab countries. The weights are on Hugging Face with vLLM and SGLang deployment support, and a suite of 24 Arabic benchmarks is being released with the code to run it.

Ashish Koshy, CEO of Inception42, puts it in one line: “Building an Arabic LLM is not a translation exercise.” Dr. Najwa Aaraj, CEO of TII, says that “sovereign AI capability must reflect the language used in daily life.” We agree with both, and our stack is built on that premise.

What has actually changed

  • Open weights make on-premises Arabic AI a serious option. When the weights can be downloaded, the model can run in your data center or in an air-gapped environment with no call to an external API. That is what makes Arabic AI genuinely auditable.
  • Dialect is now an explicit training target. A model dedicated to Emirati Arabic and speech models covering many dialects mean the question is no longer “does it understand Arabic?” but “does it understand your customers’ Arabic?”.
  • Benchmarks are public, but still the developer’s own. Figures such as 84.83% on Alyah, or the 7B Jais 3 average rising to 79.6 from 56.4 for its base model, are reported by the developer on the developer’s benchmarks. Publishing the data and code is welcome, but it is no substitute for a test on your data.

Four questions before you buy

  1. Are the weights available, and under which licence? A model offered only through a chat platform cannot be deployed air-gapped. Ask for the licence in writing, and check that it allows commercial use and fine-tuning.
  2. Which dialect was it trained on, and which dialect do your customers write in? A model that excels at Emirati Arabic is not necessarily excellent at Egyptian or Levantine. Make dialect an evaluation item, not an assumption.
  3. What size can your hardware actually run? A family from 900 million to 375 billion parameters lets you match the model to your compute budget. The biggest model is not always the best one if it has to run on limited GPUs with thousands of concurrent users.
  4. How does it perform on your questions? Take fifty real questions from your support channel, and the documents that answer them, and ask each model to answer with citations. Let the people who handle those questions every day grade the answers. This is what we do in every deployment before a single user sees it.

Where we stand

Every new open Arabic model widens what we can deploy inside your premises, with the same security controls and the same evaluation suite. Arabic is the native language of our stack, not a translation target, and the choice between one model and another is settled by numbers on your data, not by a leaderboard.

Sources

Related pages