Arabic AI for business: models, dialects and how to test
A demo in clean Arabic proves little. Your customers write in dialect, mix in English, and sometimes type Arabic in Latin letters. Here is what to look for, which models exist, and how to test with your own messages.
The short answer
Large global models and regional Arabic models can both work well. The right choice depends on your customers’ dialect, your task and where your data must live, so no leaderboard can decide it for you.
Test candidates on 50 to 100 of your own real messages, and score the answers with someone who speaks that dialect.
What makes Arabic harder
| Challenge | Example | What to check |
|---|---|---|
| Dialects | Egyptian, Gulf, Levantine and Maghrebi differ a lot | Does it understand the customer, and reply in a suitable register? |
| Franco-Arabic | “3ayez a3raf el se3r” | Does it read Arabic typed in Latin letters and numbers? |
| Mixed languages | Arabic sentences with English product names | Does it keep names, numbers and prices right? |
| Grammar and gender | Addressing a man or a woman | Does it avoid awkward or wrong forms? |
| Tone | Formal MSA can sound cold in a sales chat | Does it match your brand and the customer? |
| Display | Right-to-left text with numbers | Does it look correct in your app, not just the model’s output? |
| Voice | Accents and background noise | Test transcription separately if you use voice notes |
Which models to know
- Global models such as GPT, Claude and Gemini: strong general ability and tools, and they work across languages
- Falcon-H1 Arabic (TII, UAE): open models in 3B, 7B and 34B sizes. TII reports leading scores on the Open Arabic LLM Leaderboard, including dialect tests. These are the maker’s own claims
- ALLaM (Saudi Arabia, now under HUMAIN): built for Arabic and Saudi needs
- Jais (UAE) and Fanar (Qatar): regional Arabic-first models
- Nile-Chat: a research model for Egyptian Arabic in both Arabic and Latin scripts. Its authors report better results than several larger models on their Egyptian benchmarks
This field moves quickly, and availability and licences vary. Treat this list as a starting point rather than a ranking.
Global or regional?
- Agents that use tools and follow long instructions: test the global models first, as they usually have mature tooling
- Data that must stay in your country or on your servers: open regional models can be self-hosted, but you take on hosting, updates and security
- A narrow dialect-heavy task: include a regional or dialect-specific model in the test
- Very high volume: compare total running cost, not only the price per token
How to run a fair test
- Collect 50 to 100 real messages, with personal details removed, across your main dialects
- Include hard ones: Franco-Arabic, spelling mistakes, voice note transcripts, and angry customers
- Write what a good answer must contain, before you run anything
- Give every model the same instructions and the same company information
- Have native speakers score accuracy, tone and whether it knew to hand over, without seeing which model wrote what
- Record the cost and response time for each model
- Test again whenever you change the model or its instructions
Sources
- TII: Falcon-H1 Arabic launch (via HPCwire, January 2026)
- Nile-Chat: Egyptian language models for Arabic and Latin scripts (arXiv)
Rules change. Check the official text, or ask your accountant or lawyer, before you rely on a figure.
Build an agent that speaks like your customers
keyt builds and tests AI agents in Arabic, English and French, using real messages from your market.