BroadRouter

Gulf, Egyptian and Levantine Arabic: how LLMs handle dialects

By the BroadNet Inference team · 6 min read · Published

LLMs handle Arabic dialects unevenly: they usually understand Egyptian and Gulf Arabic better than they write them, and quality drops for rarer dialects. Test understanding and writing separately on your own messages. In BroadNet's own tests, Qwen3.8 27B understood Gulf and Egyptian dialects and wrote fluent Arabic. It costs $0.175 per 1M input tokens.

LLMs usually understand Gulf and Egyptian Arabic better than they write them, and do worse on rarer dialects. A customer in Riyadh, one in Cairo and one in Beirut can all ask “what do you want?” in three different sentences, and a model that handles Modern Standard Arabic (MSA) well can still misread them. Test understanding and writing separately.

This post covers what changes between the big dialect groups, where LLMs tend to fail, and how to test a model before you trust it with real customers.

How different are the dialects, really?

Very different in everyday words, much closer in formal vocabulary. Research corpora show the scale: the MADAR corpus translated the same travel sentences into the dialects of 25 Arab cities. The table shows a few common phrases. The MSA column is what a textbook would print; the others are typical spoken and texted forms. Spelling varies from person to person, which is part of the problem.

Meaning MSA Gulf (Saudi / Emirati) Egyptian Levantine
What do you want? ماذا تريد؟ وش تبي؟ / شو تبي؟ عايز إيه؟ شو بدّك؟
How are you? كيف حالك؟ شلونك؟ إزيك؟ كيفك؟
now الآن الحين دلوقتي هلّق
I don’t know لا أعرف ما أدري مش عارف ما بعرف
very / a lot جداً وايد أوي كتير

Three patterns matter for LLMs:

  • Function words change. Question words, negation and “want” differ completely between dialects. These are short, frequent words that decide the meaning of the sentence, so misreading one flips the intent.
  • Spelling is not standard. Dialects are mostly spoken, so people write them the way they sound. The same word can appear with different letters, with or without hamza, or in Latin script (Arabizi).
  • Code-switching is normal. Gulf business messages mix in English product names; Levantine and Maghrebi speakers mix in French. A model needs to cope with both scripts in one sentence.

Where do LLMs fail on Arabic dialects?

Models learn from the text they see, and far less public text exists for smaller dialects such as Emirati or Sudanese than for Modern Standard Arabic. In practice that shows up as a few recurring errors:

  1. Reading a dialect word as a different MSA word. Gulf “أبي” (I want) looks like MSA “أبي” (my father). Context usually resolves it, but small models sometimes do not.
  2. Answering the wrong intent. A short Egyptian complaint such as “الطلب لسه موصلش” (the order still hasn’t arrived) needs the model to know that “لسه” means “still” and “موصلش” is a negated verb.
  3. Writing a mixed dialect. Asked to reply “in Gulf Arabic”, weaker models produce a mix of Egyptian, Levantine and MSA that no native speaker would write. This is more common than misunderstanding.
  4. Over-correcting. Some models “fix” the customer’s dialect into MSA when you asked them to quote it, which breaks extraction tasks.

The general rule: models understand dialects better than they write them. Plan your product around that.

What has BroadNet tested on Arabic dialects?

BroadNet publishes its own Arabic check on its pricing page. For Qwen3.8 27B, it reports:

  • fluent Arabic writing,
  • Gulf and Egyptian dialects understood,
  • accurate Arabic–English translation.

That is what BroadNet states, and nothing more. It does not publish a Levantine, Iraqi or Maghrebi result, and it does not publish a dialect-writing score. If Levantine customers matter to you, test them yourself with the method below.

Qwen3.8 27B takes text and images, supports tools and structured output, reasons before answering and has a 262,144-token context. Per 1M tokens it costs $0.175 for input and $2.17 for output. Compare the other chat models on the pricing page, and see choosing an LLM API for Arabic for the wider checklist that also covers writing, translation and documents.

How do you prompt an LLM for a specific dialect?

Say which dialect you expect, which one you want back and how formal to be. A system prompt that works well for a Saudi retail inbox:

import os
from openai import OpenAI

client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])

system = (
    "You are a customer-service assistant for a Saudi online store. "
    "Customers write in Gulf Arabic (often Saudi), MSA, English or Arabizi. "
    "Understand all of them. Reply in simple, polite Modern Standard Arabic, "
    "in two or three short sentences. If the customer writes in English, reply in English. "
    "Never invent order details."
)

response = client.chat.completions.create(
    model="broadnet/qwen3.8-27b",
    max_tokens=2000,  # the model reasons first; leave room for the answer
    messages=[
        {"role": "system", "content": system},
        {"role": "user", "content": "السلام عليكم، طلبي للحين ما وصل، وش السالفة؟"},
    ],
)
print(response.choices[0].message.content)

The customer message is Gulf Arabic: “Hello, my order still hasn’t arrived, what’s going on?” The words “للحين” (still) and “وش السالفة” (what’s the story / what’s going on) are Gulf forms. A good reply acknowledges the delay in MSA and asks for the order number.

A few more prompt habits that help:

  • Name the country, not only the region. “Gulf Arabic” covers Saudi, Emirati, Kuwaiti, Qatari, Bahraini and Omani speech, which differ. If you know the market, say it.
  • Ask for a dialect label in extraction tasks. For routing or analytics, have the model return {"dialect": "gulf" | "egyptian" | "levantine" | "msa" | "other", "intent": ...} using structured output. It is a cheap way to learn who your customers are.
  • Keep the instructions in one language. English instructions with Arabic content work well on most models. Arabic instructions also work; mixing both in one system prompt tends to blur the register. More on this in Arabic prompt engineering.

Should you reply in dialect at all?

For most businesses, no. Simple MSA (sometimes called “white” or neutral Arabic) is understood by every Arabic reader, reads as professional and avoids the risk of a reply that sounds like the wrong country. Dialect replies make sense for informal brands, marketing copy written for one market, and voice products where MSA sounds stiff. If you go that way, have native speakers of that exact dialect review a sample of outputs every week.

How do you test dialect handling in a day?

  1. Collect real messages. Take 30 to 50 anonymised messages per dialect from your inbox, chat logs or transcribed voice notes. Remove names, phone numbers and order numbers.
  2. Label the intent. For each message, write the correct intent and any facts the model should extract (order number, product, date).
  3. Test understanding. Ask each candidate model to return the intent and fields as JSON. Score exact matches. This is the number that decides whether routing works.
  4. Test writing. For ten messages per dialect, ask for a reply. Have a native speaker grade each reply 1–3 for correctness, tone and naturalness. Do not grade dialect you do not speak.
  5. Check cost and speed. Record usage.prompt_tokens, usage.completion_tokens and latency per call. Dialect text can tokenise differently from MSA, so measure rather than guess; the free Arabic token counter gives a quick first estimate.
  6. Repeat for each model. Run the same set on Qwen3.8 27B, DeepSeek V4.1 Flash and GLM 5.3. All three use the same OpenAI-compatible API, so only the model id changes.

Qwen3.8 27B, DeepSeek V4.1 Flash and GLM 5.3 work after a top-up; among chat models, the free $5 starting credit only covers BroadNet Gemma 4 VL 26B, which is worth adding as a low-cost baseline. A useful outcome is often a split setup: one model for understanding and routing, with replies drafted in MSA.

What about dialect in voice notes?

If your customers send voice notes, dialect handling starts earlier, in speech to text. A transcription model can mishear a dialect word before the LLM ever sees it. BroadNet reports a dialect word error rate of 23.4 for BroadNet Transcribe Arabic, against 3.8 on FLEURS read speech; see Arabic WER explained for what those numbers mean. Test the whole chain, voice note to reply, not each piece alone.

Questions

Should my chatbot reply in the customer's dialect?

Usually not by default. Simple Modern Standard Arabic is understood everywhere and looks professional. Use dialect replies only where your brand voice calls for it, and only after native speakers have reviewed the output.

Which BroadNet model did BroadNet test on Arabic dialects?

Qwen3.8 27B. BroadNet reports fluent Arabic writing, understanding of Gulf and Egyptian dialects, and accurate Arabic–English translation. It has not published a Levantine or Maghrebi result.

How many test messages do I need per dialect?

Thirty to fifty real messages per dialect is enough to see clear differences between models. Add more for the dialect most of your customers use.

Do dialect messages cost more tokens?

Possibly. Token counts depend on the tokenizer and spelling, and dialect text often has non-standard spelling. Paste samples into a token counter and confirm with prompt_tokens on your own messages rather than assuming.

Can an LLM understand Egyptian Arabic?

Often, but check it on your own messages. BroadNet reports that Qwen3.8 27B understood Egyptian and Gulf Arabic in its tests; other models vary. Writing dialect is usually less reliable than reading it, so score intent accuracy and reply quality separately on 30 to 50 real messages.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity