Jais 3 and New Falcon Models Give UAE Founders an Arabic-First AI Toolkit

DUBAI — For years, founders building products for Arabic speakers have faced the same problem: AI tools that handle English well often stumble on Gulf dialects, code-switching and Arabic documents. In October, two Abu Dhabi research groups released a batch of models aimed squarely at that gap, and the Jais 3 family can be downloaded and run by any startup.

What Launched

Jais 3 (October 15). The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence and Inception42, G42’s agentic AI company, introduced Jais 3, a fleet of six open language models ranging from 900 million to 375 billion parameters, plus a speech recognition model called Jais Speech. According to the release:

• The models are built on K2 Horizon, IFM’s fully open-source foundation models released in September, and were developed in collaboration with Cerebras Systems.
• They are trained on Modern Standard Arabic and regional dialects “from the Gulf to North Africa,” and all six share a context window of 512,000 tokens, which IFM calls the longest of any Arabic-native model.
• Connected to software tools, the models can plan multi-step tasks, search the web, and read, edit and produce Arabic documents, spreadsheets, presentations and websites.
• Jais Speech transcribes Modern Standard Arabic and dialects spoken across 19 Arab countries, follows speakers who switch into English or French mid-sentence, and can identify the dialect being spoken.
• The models and Jais Speech are available on Hugging Face, with deployment support for the vLLM and SGLang serving frameworks. Each model ships with five training checkpoints, and IFM is releasing the Arabic benchmarks it used along with the code to run them.

Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic (October 6). The Technology Innovation Institute (TII), the applied research arm of Abu Dhabi’s Advanced Technology Research Council, launched three models:

• Falcon-Emirati, a 7-billion-parameter model built for Emirati Arabic vocabulary, idioms and cultural context. TII says it will be available through a Falcon Chat platform.
• Falcon-ASR, a 1.6-billion-parameter speech-to-text model covering Emirati and Modern Standard Arabic, English, French, Spanish and Portuguese. TII says it records when each word is spoken, which supports subtitling and searchable recordings.
• Falcon-OCR-Arabic, a 270-million-parameter model that extracts Arabic text and structured content, including tables, from images and documents.

What the Labs Claim, and How to Read It

Both groups publish strong benchmark results, and both ran the tests themselves. IFM and Inception42 say the largest Jais 3 model matches the best open models on Arabic understanding and leads them on agentic tasks, and that Jais Speech had the lowest average error rates of the models compared across six public test sets. TII says Falcon-Emirati scored 84.83% on Alyah, a native Emirati Arabic benchmark, ahead of every Arabic and multilingual open-source model it evaluated.

On its Falcon-ASR and Falcon-OCR-Arabic page, TII reports that the OCR model reached 81.9% overall Arabic text accuracy across 11,974 samples in 12 document categories, second of 17 models it tested and first in categories including invoices, receipts and official documents.

Self-reported benchmarks are a useful starting point, not a guarantee. The practical test is how a model performs on your own customers’ messages, calls and paperwork.

The labs frame the releases in terms of national capability. “Sovereign AI capability must reflect the language used in daily life,” said Dr. Najwa Aaraj, TII’s chief executive. Inception42’s Dr. Larry Murray pitched Jais 3 at companies: “With open models across a range of sizes, organisations can match these capabilities to their computing budgets and deployment needs.”

Where Founders Can Use Them

The labs themselves point to applications that map onto common startup products. TII lists multilingual customer service, meeting and interview transcription, subtitling, accessibility tools and voice-enabled services. For OCR, it cites document digitisation, searchable archives and automated information extraction. IFM highlights writing and agent workflows in Arabic.

Our editorial suggestions for Dubai founders:

• Start small. IFM says its smallest Jais 3 model is compact enough to run on a phone. Smaller models are cheaper to test and host, so try them first and step up only if quality demands it.
• Build a dialect test set. Collect anonymised, consented samples of the Arabic your users actually write and speak, including Emirati and other Gulf dialects and English code-switching, and compare models on that.
• Look at the paperwork. Invoices, receipts and official forms come up constantly in UAE back offices. An Arabic OCR model that handles tables could cut manual data entry for fintech, logistics and accounting tools.
• Check licences and data rules before shipping. Read each model’s licence on its download page, and make sure your hosting setup meets the data protection requirements that apply to your customers.

The broader signal is that Arabic is no longer an afterthought in open AI. For a startup in Dubai serving the region, that lowers the cost of building a product that speaks to customers in their own words.

Model details and benchmark figures are as published by IFM, Inception42 and TII as of October 28, 2026, and have not been independently tested by Entrepreneurs Of Dubai.