BroadRouter

Choosing an LLM API for Arabic: what to test

By the BroadNet Inference team · 7 min read · Published

Choose an LLM for Arabic by testing it on your own tasks: understanding Gulf, Egyptian and Levantine dialects, writing fluent Modern Standard Arabic, translating Arabic and English, and reading Arabic documents. In BroadNet's tests, Qwen3.8 27B wrote fluent Arabic and understood Gulf and Egyptian dialects. It costs $0.175 per 1M input tokens.

To choose an LLM API for Arabic, test each candidate on the jobs your product does: understanding your users’ dialects, writing natural Modern Standard Arabic, translating, and reading Arabic documents. Score each job separately with a native speaker. “Supports Arabic” can mean anything from fluent Gulf dialect to stiff, translated-sounding text.

Which Arabic jobs should you test?

Job Example What to check
Understand dialect A Saudi customer types in Khaleeji Does it get the intent, not just the words?
Write Arabic A reply in clear Modern Standard Arabic Fluency, grammar, natural tone, right register
Translate Arabic ↔ English contracts or support tickets Accuracy of names, numbers and legal terms
Read documents Photo of an Arabic invoice Totals, dates and vendor extracted correctly
Hear speech Voice notes, calls That’s a speech-to-text model, not an LLM; see below

A model can be strong on one and weak on another. Score each job separately.

What do public Arabic leaderboards tell you?

Several public leaderboards rank models in Arabic. The Open Arabic LLM Leaderboard on Hugging Face is one. BALSAM is a community benchmark with 78 tasks in 14 categories and a platform for blind evaluation; its authors write that LLM performance in Arabic lags behind the progress seen in English.

Use these to build a shortlist, not to make the final choice. Benchmarks are mostly multiple-choice questions and standard tasks. They rarely test Gulf customer messages, your brand’s tone, or a photo of a crumpled receipt. Those are the things your users notice.

Which LLM APIs support Arabic?

Almost every general chat API accepts Arabic text, so the useful question is how well it handles your dialects and whether you can call the Arabic-first models at all. We checked the four Arabic-first families people ask about on 14 November 2026. None had a per-token API price published by its maker:

Model family Who makes it How you can run it (checked 14 November 2026) Price
Jais 2 (8B, 70B chat) MBZUAI, Inception and Cerebras Open weights on Hugging Face under Apache 2.0, 8,192-token context. The 70B model card says no Inference Provider deploys it. The older Jais-30b-chat was removed from GitHub Models on 31 October 2025. None published; you pay for your own GPUs
Falcon-H1-Arabic (3B, 7B, 34B) Technology Innovation Institute (TII) Announced 5 January 2026 with 128K (3B) or 256K (7B, 34B) context. The launch post links a web chat demo; it names no hosted API and no weights download. None published
ALLaM (allam-1-13b-instruct) Saudi Data and Artificial Intelligence Authority (SDAIA); HUMAIN publishes the 7B preview On IBM watsonx.ai as a deploy-on-demand model with a 4,096-token window (input plus output). ALLaM-7B-Instruct-preview is open on Hugging Face under Apache 2.0. watsonx.ai bills deploy-on-demand models per hour, not per token
Fanar Qatar Computing Research Institute (QCRI) Fanar says it “provides an API for developers”; access is by request form at api.fanar.qa None published

So if you want pay-per-token Arabic through an OpenAI-compatible API, you will usually use a strong multilingual model, such as the ones below, and test it on your dialects. If you need a specific Arabic-first model, plan to self-host it, rent a dedicated deployment, or request Fanar API access; the self-hosting vs API cost post shows the break-even math.

Which BroadNet models can handle Arabic?

BroadNet publishes its own Arabic check on its pricing page. For Qwen3.8 27B it reports fluent Arabic writing, Gulf and Egyptian dialects understood, and accurate Arabic–English translation. The model takes text and images, supports tool calling and structured output, and has a 262,144-token context.

Other options on the same API:

Model Input / output per 1M tokens Context Images
Qwen3.8 27B $0.175 / $2.17 262,144 Yes
DeepSeek V4.1 Flash $0.175 / $0.52 1,048,576 Yes
BroadNet Gemma 4 VL 26B $0.10 / $0.22 32,768 Yes

DeepSeek V4.1 Flash is also available unlimited on a flat plan ($60 a month per session), which suits high-volume Arabic chat and agents. The DeepSeek V4.1 Flash vs Qwen3.8 27B comparison puts the two main candidates side by side. Gemma 4 VL 26B is the cheapest BroadNet model per token and works on the free starting credit.

How do you build an Arabic test set?

  1. Collect 50 real inputs from your product: chats, tickets, documents. Include at least 10 in each dialect your users write. For a dialect-only test, with an intent-labelling method, see how LLMs handle Gulf, Egyptian and Levantine Arabic.
  2. Write the ideal answer for 20 of them, or have a native speaker do it.
  3. Run every candidate model on all 50 with the same system prompt. State the dialect and register you want in the prompt: “Reply in clear Modern Standard Arabic, friendly, short.”
  4. Blind-rate the outputs 1 to 5 for correctness and naturalness. Have a native speaker do the rating.
  5. Check cost on real text. Read prompt_tokens and completion_tokens from the usage block rather than estimating from word counts.

Because the API is OpenAI-compatible, switching models is a one-line change of the model field. See the OpenAI-compatible migration checklist.

How should you score the outputs?

Criterion Score 1 Score 5
Intent understood Misreads the dialect Correct, including slang
Grammar and fluency Errors or broken agreement Reads like a native writer
Register Wrong dialect or too formal Exactly what the prompt asked
Facts, names, numbers Changed or invented Exact
Script Latin transliteration or mixed Arabic script throughout

Average per job, not overall. A model that scores 4.5 on writing and 2 on dialect is a different choice from one that scores 3.5 on both.

Does the system prompt change the result?

Yes, enough to change the ranking. Use the same system prompt for every candidate, and state the dialect you expect, the register you want back, and how to treat numbers and English names. Arabic prompt engineering has a complete template. Keep that prompt identical across calls so it can be read from cache on models with a cached input price.

Why does Arabic cost more tokens than English?

Arabic text often splits into more tokens per word than English, and the ratio differs between models. Two models with the same price per token can have different costs per Arabic message. Paste a sample of your real text into the Arabic token counter, then confirm with prompt_tokens from a live request on each candidate and project the monthly bill with the LLM API cost calculator. How to measure your own ratio is in tokens per word in Arabic vs English. For what this does to the bill on each model and provider, see our real cost by language dataset, which prices 1M characters of Arabic, its dialects and other languages with each model’s own tokenizer.

Can an LLM read Arabic invoices and images?

For invoices, receipts and forms, send the photo to a vision-language model and ask for JSON. The Arabic invoice guide has working code, and Arabic OCR vs vision LLMs covers when classic OCR is still the better tool. Check numbers carefully: Eastern Arabic numerals (٠١٢٣) and Western numerals (0123) both appear on Gulf receipts.

What about Arabic voice notes and calls?

For voice notes and calls, transcribe first with an Arabic speech-to-text model, then send the text to the LLM. BroadNet Transcribe Arabic costs $0.0042 per audio minute. How to choose an Arabic speech-to-text API.

Why do some Arabic answers come back empty?

DeepSeek V4.1 Flash and Qwen3.8 27B reason before they answer, and those reasoning tokens count toward max_tokens. Arabic answers can be long, so leave room, or the visible answer comes back empty. Why that happens and how to fix it.

What mistakes should you avoid?

  • Rating with non-native speakers. Fluency problems are invisible to learners.
  • Mixing dialects in one score. Keep Gulf, Egyptian and Levantine results apart.
  • Trusting translated test sets. Text translated from English doesn’t read like real Arabic users.
  • Forgetting right-to-left display. Mixed Arabic and English lines can render out of order in your UI. Test the display, not only the text.

Where BroadNet Inference fits

BroadNet offers an Arabic-tested chat model, two cheaper or unlimited alternatives, and Arabic speech to text on one OpenAI-compatible key, with no content retained. It does not publish Arabic benchmark scores for its chat models beyond its own tests, so run the test set above before you commit. If a model elsewhere wins on your dialect, that result matters more than any catalogue.

Questions

Which BroadNet model is best for Arabic?

BroadNet tested Qwen3.8 27B in Arabic and reports fluent writing, Gulf and Egyptian dialects understood, and accurate Arabic–English translation. Test it, and DeepSeek V4.1 Flash, on your own prompts.

What is the best LLM for Arabic?

It depends on the job and the dialect. Use public leaderboards such as the Open Arabic LLM Leaderboard or BALSAM to shortlist, then rate the shortlist on 50 of your own inputs with a native speaker. A model that leads a benchmark can still sound stiff in your dialect.

Can an LLM read Arabic documents and images?

Vision models such as Qwen3.8 27B, DeepSeek V4.1 Flash and BroadNet Gemma 4 VL 26B accept images, so they can read photos of Arabic invoices, receipts and forms and return JSON.

Which LLM APIs support Arabic?

Most general chat APIs accept Arabic, but quality varies by dialect. Arabic-first models such as Jais 2, Falcon-H1-Arabic, ALLaM and Fanar had no maker-published per-token API price on 14 November 2026. Jais 2 is open weights to self-host, ALLaM runs on IBM watsonx.ai as an hourly dedicated deployment, and Fanar gives API access by request.

Why does Arabic text cost more tokens than English?

Tokenizers are often trained on more English than Arabic, so Arabic words tend to split into more tokens. The effect varies by model, so compare prompt_tokens on the same Arabic text for each candidate.

Can an LLM reply in Gulf or Egyptian dialect?

Many can, but quality varies more than in Modern Standard Arabic. Ask for the dialect explicitly in the system prompt and have a native speaker rate the replies.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity