Key facts
| Arabic chat | Handled by multilingual chat models |
| Arabic embeddings | plugsky-embed-multilingual for retrieval and RAG |
| API | OpenAI-compatible with UTF-8 Arabic input and output |
| Model choice | 30+ models; benchmark per dialect and task |
| RAG | Arabic retrieval supported via multilingual embeddings |
| Free tier | 2 free AI models (plugsky-micro, plugsky-lite), no card |
| Trial | 14-day full-access trial available |
| Product status | Live |
TL;DR
- Arabic chat and generation work through the standard OpenAI-compatible API.
- plugsky-embed-multilingual is the embedding model for Arabic retrieval.
- Quality differs by model and dialect — test MSA and your target dialect separately.
- Build Arabic RAG on multilingual embeddings plus a chat model.
- Begin on the free plan; use the full-access trial to compare stronger models.
How it works, step by step
- Collect a small Arabic evaluation set from your real use case, including dialect samples.
- Benchmark two or three chat models on the same prompts and score accuracy, fluency and formatting.
- Index your Arabic documents with plugsky-embed-multilingual and test retrieval precision.
- Check right-to-left rendering and mixed Arabic/English text in your UI end to end.
- Tune prompts for Arabic response style rather than translating English prompts literally.
- Monitor quality after launch with Arabic-language feedback and periodic eval reruns.
Original data
Try it yourself
Open the embedding model comparison →
What Arabic support looks like
Arabic works across the stack. Chat models handle Arabic prompts and produce Arabic responses, including mixed Arabic-and-English conversations common in Gulf enterprises. For retrieval, plugsky-embed-multilingual maps Arabic text into the same vector space as your other languages, so one index can serve multilingual search. The API itself is language-agnostic: text is UTF-8, and no special endpoint or parameter is needed. The interesting question is not whether Arabic works, but which model works best for your dialect and task.
Where quality varies
Be sceptical of blanket claims about Arabic performance:
- Dialect: Modern Standard Arabic, Gulf, Egyptian and Levantine differ; a model strong on MSA may be weaker on dialect-heavy social text.
- Task: translation, summarisation, extraction and long-form generation stress different capabilities.
- Script handling: diacritics, numerals and mixed-script text can trip tokenisation and formatting.
- Domain vocabulary: legal, financial and medical Arabic benefit from retrieval augmentation rather than relying on the model alone.
Run a small, labelled evaluation on your own data; it is the only benchmark that predicts your production quality.
What we do and what we do not do
What we do: provide multilingual chat models and a multilingual embedding model behind one API, and support Arabic RAG with the same building blocks other languages use. What we do not do: publish a single Arabic leaderboard winner — there is no one model that is best for every dialect and task, and quality moves as models update. Pick with your own evals, revisit periodically, and keep the model name in configuration so you can switch without a code change. See the model catalogue for current options.
Honest comparison
| Need | Plugsky | Monolingual English API | Self-hosted model |
|---|---|---|---|
| Arabic chat | Multilingual models, live | Often weak | Depends on model choice |
| Arabic embeddings | plugsky-embed-multilingual | Usually English-first | You deploy and tune |
| Dialect coverage | Benchmark per dialect | Limited | Depends on checkpoint |
| Arabic RAG | Multilingual embeddings plus chat models | Requires workarounds | You build the stack |
| Operational effort | API call | API call | GPUs and serving |
| Best fit | Arabic apps and assistants | English-only products | Strict offline needs |
Frequently asked questions
Is Arabic supported on the free plan?
The free plan includes plugsky-micro and plugsky-lite. Both are usable for Arabic prototyping; use the 14-day full-access trial to compare stronger models on your own data.
Which embedding model should Arabic RAG use?
plugsky-embed-multilingual — it places Arabic and other languages in a shared vector space, so one index can serve mixed-language retrieval.
Do I need a special endpoint for Arabic?
No. Send Arabic as UTF-8 text to the same OpenAI-compatible chat and embeddings endpoints used for any other language.
How do I choose between models for Arabic?
Build a small labelled evaluation set from your real workload, including dialect samples, and score candidates on accuracy, fluency, formatting and cost.
Can it handle mixed Arabic and English?
Yes, code-switching is common in business text and multilingual models generally handle it. Test it explicitly, because formatting rules for mixed text can vary.
Is Arabic OCR or speech supported?
Those depend on audio and vision endpoints, which are coming soon. For text chat, generation and embeddings, Arabic is supported today.
How do I improve Arabic quality without changing models?
Use RAG with multilingual embeddings to supply domain vocabulary and facts, tune prompts for Arabic response style, and keep evaluation running after launch.