Arabic and AI in 2026: From Egypt's Karnak to the Region's Models

2026 was a notable year for Arabic in AI: Egypt launched its national Karnak model and the region built models of its own. Dialects are still a real challenge, though. Here's what that means for businesses.

20 Oct 20265 min read

A winding line of light like the Nile, branching out at the top like a delta

For years Arabic was a second-tier language in AI. Models understood it, but clearly worse than English, and struggled even more with dialects. In 2026 the region moved seriously to change that, and Egypt played a visible part.

Egypt's AI strategy for 2025–2030

In January 2025 Egypt launched the second edition of its National AI Strategy, covering 2025 to 2030. Its headline targets include:

Raising the ICT sector's share of GDP to 7.7% by 2030.
Reaching more than 250 AI startups and specialized companies.
Building a workforce of 30,000 AI professionals.

The strategy rests on six pillars: governance, infrastructure, technology, data, the business ecosystem and talent. One of its key commitments is a national Arabic foundation language model, followed by specialized models for sectors such as agriculture, health and law.

Karnak: Egypt's language model

On February 11, 2026, at the opening of Ai Everything MEA Egypt in Cairo, Egypt's Ministry of Communications and Information Technology announced Karnak, the country's national Arabic language model. Egypt's Information Technology Industry Development Agency (ITIDA) describes it as the highest-ranking Arabic model in two size categories, without disclosing its exact parameter count.

Applications announced as running on Karnak include:

SIA: an interactive tutor for Arabic and Egyptian history.
A legal and regulatory assistant that helps citizens and small businesses understand procedures and regulations.

In April 2026, reports described exploratory talks between the Egyptian government and Microsoft about integrating Karnak with Microsoft's cloud services in Egypt, with no final agreement announced so far. In September 2026, the Ministry of Communications announced the Karnak Chat assistant and the Loghat app as part of its "digital sovereignty" drive. Before that, in August 2026, the free BelMasry app launched, offering speech-to-text and text-to-speech in Arabic, Egyptian Arabic and other languages.

The region's models: a real Arabic race

Egypt isn't alone. Notable Arabic models in the region include:

Jais 2 from the UAE (Inception, a G42 company, with Mohamed bin Zayed University of AI and Cerebras): released in December 2025 with open weights, covering Modern Standard Arabic and regional dialects.
ALLaM and HUMAIN Chat from Saudi Arabia: an Arabic assistant built on the ALLaM model, hosted entirely in the Kingdom, and supporting dialects including Egyptian and Lebanese.
Falcon Arabic from Abu Dhabi's Technology Innovation Institute, followed in October 2026 by Falcon-Emirati for the Emirati dialect, alongside speech-to-text and Arabic OCR models.
Fanar 2.0 from Qatar (Qatar Computing Research Institute): released in December 2025, adding multimodal capabilities.

The uncomfortable truth: dialects are still hard

Despite the progress, research shows a clear gap. The DialectalArabicMMLU study (October 2025), which tested 19 open models, found average accuracy of:

62.8% in English,
51.9% in Modern Standard Arabic,
47.7% across Arabic dialects, with Egyptian Arabic averaging 48.9%.

The best result on Egyptian Arabic came from a model built specifically for it (Nile-Chat), at 61.9%. The study also found models were poor at identifying which dialect they were reading. The study covered small and mid-sized open models, not the latest large commercial ones, and we found no independent 2026 evaluation of the newest models on Egyptian Arabic specifically. The message is still clear: don't assume a model understands your customers the way your staff do.

What this means for businesses in Egypt

Test in your customers' real language. Before launching an AI assistant, try it on hundreds of real customer messages, with their dialect, typos and shorthand.
Match the task to the model's strengths. Formal writing and summarizing in Modern Standard Arabic are strengths of today's models. Understanding long colloquial complaints or local jokes needs more careful testing and human oversight.
Watch local and regional options. Arabic models hosted inside the region may become a real advantage, especially as Egypt's data protection law comes into force with its limits on transferring data abroad.
Don't rush to drop global models. The major models are still very strong in Modern Standard Arabic and on complex tasks. The answer is often a mix that fits each task.

The bottom line

Arabic is no longer on the margins of AI. Egypt now has a national model and a clear strategy, and the region has a genuine race for Arabic models. Dialects, Egyptian Arabic included, remain the area where businesses need careful testing before relying on AI fully.

At CODE NUTS we build systems and AI solutions that work in Arabic and English, and we test them on customers' real language. Explore our services.