Short answer. Arabic support is not one feature. Translation has explicit Arabic coverage in the model behind
plugsky-translate, Arabic is among the languages evaluated forplugsky-embed-multilingual, and Mistral documents Arabic among the languages with strong expected performance. Most Nemotron-based chat IDs do not list Arabic, and the Nano Omni model is English-only. Treat Arabic generation and OCR as best-effort unless your own evaluation proves otherwise.
Quick answers
These answers separate documented language support from best-effort behavior. Model cards are vendor claims; your evaluation is the evidence that matters.
| Buyer question | One-sentence answer |
|---|---|
| Does Plugsky support Arabic? | Partially and unevenly: translation is explicit, embeddings are evaluated on Arabic, and chat support depends on the model ID. |
| Which model is best for Arabic chat? | plugsky-mistral-medium is the best-documented route, but test MSA and your dialect before committing. |
| Do the Nemotron models support Arabic? | NVIDIA's model card does not list Arabic for Nemotron 3 Ultra, and Nano Omni is English-only, so treat those IDs as best-effort. |
| Is Gulf dialect supported? | No dialect-specific tuning is published; MSA is the safer target, and Gulf, Levantine and Egyptian are evaluation tasks. |
| Does the API handle right-to-left (RTL) text? | The API returns UTF-8 text; RTL layout and bidirectional (bidi) isolation are client-side concerns you must implement and test. |
| Are the embeddings multilingual? | plugsky-embed-multilingual is 2048-dimensional and evaluated across 34 languages including Arabic, so validate retrieval on your corpus. |
| Where is Arabic data processed? | The live GCC region is UAE me-central-1; Riyadh is Enterprise-only, so confirm the path in writing. |
| Can it OCR Arabic documents? | plugsky-parse uses an English-only primary engine, so Arabic OCR is not a supported claim; test it or use a specialist. |
What the catalogue actually supports
Model names are a poor guide to language support. The serving engine and its model card are better, and your own evaluation set is better still.
| Job | Public ID | Arabic status | What to test |
|---|---|---|---|
| Translation | plugsky-translate | Explicit: Arabic is listed for Riva Translate 4B, with published BLEU/COMET scores. | Domain terms, numbers, MSA and Gulf tone. |
| Chat and drafting | plugsky-mistral-medium | Documented: Arabic appears in Mistral's supported-language list; no Arabic benchmark is published. | MSA, dialect, long documents, tool use. |
| Chat and drafting | Nemotron-based IDs | Not listed for Nemotron 3 Ultra; Nano Omni is English-only. | Do not assume; evaluate before use. |
| Embeddings and RAG | plugsky-embed-multilingual | Evaluated among 34 languages including Arabic; Plugsky labels it best-effort multilingual. | Recall@k on your Arabic corpus. |
| OCR (optical character recognition) and documents | plugsky-parse | Primary engine is English-only Nano Omni; Arabic OCR is not a supported claim. | Avoid, or test a specialist. |
Which Plugsky models can handle Arabic?
Start with the job, then the model. For translation, plugsky-translate is built on NVIDIA Riva Translate 4B, whose model card lists Arabic and publishes Arabic BLEU/COMET scores. For retrieval, plugsky-embed-multilingual is built on Nemotron 3 Embed 1B, evaluated across 34 languages including Arabic. For chat, plugsky-mistral-medium is the clearest documented route: Mistral documents Arabic among the languages with strong expected performance. NVIDIA's card reports 24.6 sacreBLEU for English→Arabic and 40.1 for Arabic→English; treat those as vendor evaluations, not a Plugsky benchmark. The Nemotron 3 Ultra chat models do not list Arabic in NVIDIA's model card, and Nano Omni is English-only. Treat those IDs as best-effort for Arabic.
Is Modern Standard Arabic supported as well as English?
For translation, yes: Arabic is an explicit language pair for plugsky-translate, in both directions. For chat, the answer depends on the model. Mistral documents Arabic as a supported language, but neither Mistral nor Plugsky publishes an MSA-specific benchmark. The honest position is that MSA is the most likely register to work well, and you should measure it on your own prompts. Do not treat a few fluent answers, or an English-language evaluation, as evidence of consistent MSA quality across long documents, numbers and legal terminology.
How good is the model at Gulf dialects?
We found no Gulf-dialect benchmark from the model vendors in this stack, and Plugsky publishes no dialect-specific tuning claim. Gulf Arabic mixes MSA, local lexicon, English loanwords and code-switching; a model that handles MSA can still miss intent, politeness or regional meaning. If your product serves Gulf customers, build an evaluation set with real, consented examples across Saudi, Emirati, Qatari and Kuwaiti usage, plus code-switched messages. Score comprehension, tone and factual accuracy separately. Treat dialect quality as a property of your evaluation, not a line on a model card.
Can one model handle Arabic and English in the same conversation?
Usually for comprehension, but quality is asymmetric. A model may understand a mixed message and still answer in the wrong language, switch script mid-sentence, or translate an English brand name unnecessarily. For bilingual products, instruct the output language explicitly, test code-switching, and check that numbers, dates and names survive. If you need strict language control, add a post-generation language check and a fallback. Do not assume that a model trained on multilingual data will follow your language policy without evaluation.
What is the difference between translation and generation for Arabic?
Translation is a constrained transformation: the source meaning is given, and the model's job is to preserve it. Generation is open-ended: the model must decide content, structure and tone, and Arabic quality depends on its pretraining and instruction tuning. That is why plugsky-translate can be the right tool for localization while the same task in a chat model may be weaker. Use translation for content conversion, and use a chat model for drafting, summarization or analysis only after you have tested it. Do not use a translation model for advice or reasoning.
Script, rendering and retrieval
Arabic adds failure modes that an English-only test plan will not catch: direction, shaping, normalization and token density.
Does the API handle right-to-left text?
The API accepts and returns UTF-8 text, and JSON does not carry a text direction. RTL is a rendering concern in your application: set dir="rtl" and the correct lang on the container, use a font with Arabic glyphs, and test line breaking, punctuation and numbers. For mixed Arabic and Latin content, bidi isolation matters. The API will not add directional marks for you, and it will not fix a UI that renders Arabic backwards. Treat RTL as a front-end acceptance test, not an API feature.
What is bidirectional text and why does it break?
Bidirectional text mixes right-to-left and left-to-right runs, such as Arabic with English product names, URLs, phone numbers or code. The Unicode Bidirectional Algorithm orders those runs for display, but ambiguous characters and missing isolation can move punctuation or reorder numbers. In HTML, use dir attributes and, where needed, bdi or Unicode isolate controls. In plain-text exports, test how the text appears in email, PDF and spreadsheets. Bidi bugs are visual, so validate with Arabic-speaking reviewers on real devices, not only with a string comparison.
Which tokenizer issues affect Arabic?
Arabic is morphologically rich, with clitics, prefixes and suffixes that can change token counts, and diacritics that are often optional. Token counts for Arabic can differ substantially from English for the same meaning, so do not estimate context or cost from an English ratio. Normalization also matters: alef variants, ta marbuta, hamza and tatweel can make semantically identical strings look different to a retriever. For search and deduplication, apply a documented normalization policy and keep the original text. Measure tokens on your own corpus before you set limits.
How should we chunk Arabic documents for retrieval-augmented generation (RAG)?
Chunk on meaning, not on a fixed character count. Arabic paragraphs can be long, and a naive split can separate a pronoun from its referent or cut a legal clause in half. Prefer section and heading boundaries, keep a small overlap, and preserve the document's metadata. Test chunk sizes against your retrieval set; Arabic token density may push you to smaller chunks than English. For mixed Arabic-English documents, do not translate before chunking unless you have evaluated the loss. Store the original text and the normalized text side by side.
Are the embeddings multilingual?
plugsky-embed-multilingual is a 2048-dimensional embedding model with an 8,192-token context, built on NVIDIA Nemotron 3 Embed 1B. NVIDIA's model card says it was evaluated across 34 languages including Arabic, with multilingual and cross-lingual retrieval as the intended use. That is a reason to test it, not a guarantee for your dialect or domain. Plugsky labels the model best-effort multilingual. Measure recall on Arabic queries against Arabic and English documents, and compare with a dedicated Arabic embedding model if retrieval quality is business-critical. See the model page for the API details.
How do we evaluate Arabic retrieval?
Build a question set from real user language, including MSA, dialect, code-switching, typos and transliteration. Pair each question with the passages that should be retrieved, then measure recall at k, mean reciprocal rank and answer faithfulness. Test cross-lingual cases where the query is Arabic and the document is English, and the reverse. Check normalization, diacritics and numeral variants. Keep a holdout set and re-run after any embedding or chunking change. A model that retrieves well on translated Wikipedia is not evidence for your contracts, tickets or product catalogue. The RAG guide covers the wider pipeline.
Hosting, residency and dialects
Language quality and data location are separate decisions. Do not let a regional sales presence stand in for a data path.
Which GCC regions are live?
Plugsky's live GCC region is me-central-1 in the UAE. Riyadh is available on Enterprise, not as a self-serve default. That means a standard plan should be assumed to process in the UAE unless the contract says otherwise. For Saudi customers, ask for the Enterprise architecture and confirm where prompts, logs, backups and support access are located. For UAE customers, me-central-1 is the natural choice, but DIFC and ADGM entities still need their own data-protection analysis. Do not infer residency from a currency or a sales office. See data residency.
Is Riyadh available self-serve?
No. Riyadh is listed as available on Enterprise, not as a self-serve default region. If in-Kingdom processing is a requirement, ask for an Enterprise architecture review and get the data path in writing before you design around it. Confirm which services run in Riyadh, where logs and backups sit, and whether support access originates outside the Kingdom. If the answer does not meet your regulator's expectation, the honest alternatives are the UAE region, a private VPC, on-premises or air-gapped deployment. Each has a different operational cost.
What should we ask about Arabic data residency?
Ask where the prompt is processed, where the response is logged, where embeddings and backups are stored, and which support staff can access them. Region names are not enough: a UAE region does not automatically satisfy a Saudi or DIFC requirement, and a private VPC can still have remote support access. Ask for a data-flow diagram and the subprocessor list. For personal data, map the transfer mechanism and have counsel review it. For models, ask whether the serving engine can change and where the fallback runs. The residency cluster goes deeper.
How do we handle dialect-specific evaluation?
Write a small, high-quality set per dialect rather than one giant “Arabic” set. Include customer-support messages, voice-note transcripts, social comments and formal documents, because register changes the task. Have native speakers label intent, tone and acceptability, and track inter-annotator agreement. Keep a separate MSA set so you can see whether a dialect gain costs you formal accuracy. Re-run after model or prompt changes. If you cannot resource native review, say so and limit the claim: “evaluated on MSA” is more honest than “supports Arabic”.
Can we fine-tune for Arabic?
Fine-tuning is not part of the published self-serve plan features, so ask whether it is available under Enterprise and who owns the resulting weights. If it is not, domain adaptation has to come from retrieval, prompt design, terminology lists and a model with documented Arabic support. Do not assume that a fine-tune is portable, that it can be exported, or that it will keep the vendor's safety behavior. For most teams, a measured RAG pipeline and a clear evaluation set deliver more Arabic improvement than an unverified fine-tune.
What to verify in a contract
Language support is easy to over-promise. Put the commitment, the fallback and the acceptance test in writing.
- Model and engine. The public ID, the serving engine, the fallback chain and the notice you get before either changes.
- Language commitment. Which languages are supported, which are best-effort, and the acceptance criteria for each.
- Evaluation. Who owns the Arabic test set, whether you can publish results, and whether regression runs are required.
- Residency. Where prompts, logs, embeddings and backups sit, and which support staff can access them.
- Retention and training. Whether Arabic content is stored, used for training or reviewed by humans.
- Embeddings. Dimension, context limit, normalization assumptions and what happens to the index at exit.
- OCR and documents. Whether Arabic OCR is in scope, and the accuracy threshold you require.
- Deprecation. Notice and migration support if an Arabic-capable model or endpoint is retired.
- SLA by region. Whether the SLA applies in
me-central-1and Riyadh, and what the exclusions are. - Subprocessors. The model provider, hosting provider and support locations, with change notice.
What we did not verify
Plugsky does not publish an Arabic benchmark, and we did not run a systematic MSA or dialect evaluation for this page. The language lists we cite are vendor model cards, not Plugsky guarantees. We did not test Arabic OCR, long-document retrieval or code-switching at scale, and we did not verify Mistral's Arabic quality beyond its published language list. Riyadh availability, Arabic-language support terms and embedding limits are contract-led. The embedding model card also describes a smaller hosted-API limit than Plugsky's catalogue, so confirm the serving limit before you size chunks. Where a model card and the Plugsky catalogue differ, the API and the signed terms govern.
Frequently asked questions
Is Arabic officially supported by Plugsky?
It depends on the job. Translation and multilingual embeddings have documented Arabic support; chat support depends on the model, and most Nemotron chat IDs do not list Arabic.
Which model should we use for Arabic chat?
plugsky-mistral-medium is the best-documented chat route because Mistral documents Arabic. Test MSA and your dialect on your own prompts before production.
Do you support Gulf dialects?
No dialect-specific tuning is published. Gulf, Levantine and Egyptian Arabic are best-effort; evaluate with native speakers and consented examples.
Does the API add RTL formatting?
No. It returns UTF-8 text. Your application must set direction, choose an Arabic font and handle bidirectional runs.
Do the embeddings handle Arabic?
NVIDIA evaluated Nemotron 3 Embed 1B across 34 languages including Arabic. Plugsky labels the multilingual model best-effort, so measure retrieval on your corpus.
Can it translate Arabic to English?
Yes. plugsky-translate is built on Riva Translate 4B, which lists Arabic in both directions and publishes BLEU/COMET scores.
Can it OCR Arabic PDFs?
Not as a supported claim. plugsky-parse uses an English-only primary engine. Test a sample or use a specialist Arabic OCR service.
Where is Arabic data processed?
The live GCC region is me-central-1 in the UAE. Riyadh is Enterprise-only; confirm the data path, logs and support access in writing.
Is Riyadh self-serve?
No. Riyadh is an Enterprise option, not a default self-serve region. Ask for an architecture review and a written data-flow diagram.
How do we estimate Arabic token costs?
Measure tokens on your own Arabic corpus. Morphology and diacritics make English token ratios unreliable, and plan ceilings are in tokens.
Can we fine-tune an Arabic model?
Fine-tuning is not in the published self-serve features. Ask about Enterprise availability, data ownership, export and safety evaluation.
Does the model understand Arabic and English mixed together?
Often for comprehension, but output language control is unreliable. Test code-switching and add a language check if it matters.
Test Arabic on your own corpus
Start with translation and embeddings, then evaluate chat on the dialect and documents you actually use.
Start free → Arabic embeddingsSources
Primary model cards and standards we checked. Vendor language lists are claims; the API and your evaluation are the evidence.
- NVIDIA — Riva Translate 4B Instruct v2 model card (Arabic listed, BLEU/COMET) — checked 20 November 2026.
- NVIDIA — Nemotron 3 Ultra 550B model card (supported languages) — checked 20 November 2026.
- NVIDIA — Nemotron 3 Nano Omni model card (English only) — checked 20 November 2026.
- NVIDIA — Nemotron 3 Embed 1B model card (34 languages including Arabic) — checked 20 November 2026.
- Mistral — supported languages, including Arabic — checked 20 November 2026.
- Unicode — UAX #9, Unicode Bidirectional Algorithm — checked 20 November 2026.
Last checked 20 November 2026. This is a 2027 planning guide, not a guarantee of Arabic quality or residency.