Key facts
| Chat shortlist | plugsky-lite or plugsky-micro for volume; plugsky-plus or plugsky-pro for quality |
| Arabic retrieval | plugsky-embed-multilingual for cross-language embeddings and RAG |
| Endpoints | OpenAI-compatible /v1/chat/completions and /v1/embeddings |
| Context class | 128K-class windows on current models; live limits are published per model |
| Free plan | Two free models, plugsky-micro and plugsky-lite, with no card required |
| Trial | 14-day full-access trial for evaluating paid tiers |
| Evaluation | Score candidates on your own MSA, dialect and mixed-language prompts |
| Product status | Live |
TL;DR
- Start free: score plugsky-micro and plugsky-lite on your own Arabic prompts.
- Escalate tone-sensitive or long Arabic writing to plugsky-plus or plugsky-pro.
- Use plugsky-embed-multilingual for Arabic semantic search and RAG.
- Dialect and code-switching quality varies — always evaluate before committing.
- Keep an Arabic eval set in CI so model updates never regress silently.
How it works, step by step
- Collect a representative Arabic eval set: MSA, your target dialect and mixed Arabic-English prompts.
- Score the free models, plugsky-micro and plugsky-lite, on that set first.
- Add plugsky-plus and plugsky-pro to the comparison for summarisation and tone-sensitive work.
- Test retrieval separately with plugsky-embed-multilingual on real documents and queries.
- Pick a default model per workload and define escalation rules for hard or long inputs.
- Pin the eval set in CI and re-run it whenever the live model card changes.
Original data
Try it yourself
Which Plugsky models handle Arabic best
Arabic quality depends less on a single benchmark than on the task type. Short chat, intent classification and tagging are usually fine on the free tiers: plugsky-lite for speed and plugsky-micro where a little more reasoning helps. Tone-sensitive replies, structured summarisation and long reports usually improve when you move up to the paid workhorse tiers such as plugsky-plus or plugsky-pro.
Check the live model card at /models before wiring constraints into your app: context window, capability flags and the engine currently serving each model can change.
Arabic retrieval: embeddings matter as much as chat
Most Arabic products are retrieval problems wearing a chat costume. If your RAG pipeline embeds Arabic documents with an English-only model, answer quality drops before the chat model is even involved. Use plugsky-embed-multilingual for mixed Arabic and English corpora, and keep a same-language baseline (plugsky-embed) to measure whether multilingual retrieval actually helps on your data.
- Normalise text before embedding: strip tatweel, unify alef forms and handle diacritics deliberately.
- Chunk on sentence and paragraph boundaries rather than fixed character counts for Arabic prose.
- Measure retrieval recall separately from answer quality so you know which stage to fix.
- Re-embed the whole index when you change embedding models, and version the index alongside the model name.
A practical Arabic evaluation plan
Treat model selection as an evaluation project, not a purchase decision. Build 50-100 real prompts across chat, extraction, summarisation and retrieval, with a scoring rubric your team agrees on. Run every candidate through the same set, including the free tiers, and record failures with the exact dialect and input length.
Then encode the outcome as routing rules: routine Arabic traffic stays on the cheapest tier that passes, long or high-stakes requests escalate, and retrieval uses the embedding model that won. Self-serve plans are flat monthly with fair-use usage, so align the routing policy with the live pricing page rather than per-token estimates.
Honest comparison
| Workload | Free models (micro, lite) | Paid chat (plus, pro) | Embeddings and long-context |
|---|---|---|---|
| High-volume Arabic chat | Good starting point, no card | Better nuance and consistency | Not needed |
| Long Arabic documents | Short inputs only | Good for most reports | Use a long-context tier for multi-document jobs |
| Arabic semantic search | Not applicable | Not applicable | plugsky-embed-multilingual |
| Dialect and code-switching | Evaluate before shipping | Usually stronger | Measure retrieval recall separately |
| Selection effort | Free plan plus 14-day full-access trial | Same eval set | Same eval set |
Frequently asked questions
Which Plugsky model is best for Arabic?
There is no single winner: use plugsky-lite or plugsky-micro for volume, plugsky-plus or plugsky-pro for tone-sensitive work, and plugsky-embed-multilingual for retrieval. Score the candidates on your own Arabic prompts before deciding.
Does Plugsky support right-to-left Arabic text?
Yes — chat models accept Arabic text through the OpenAI-compatible chat endpoint. Your front-end is responsible for RTL rendering and bidi handling; the API treats Arabic as ordinary UTF-8 input.
Which embedding model should I use for Arabic RAG?
plugsky-embed-multilingual is the cross-language option for corpora that mix Arabic and English. Keep plugsky-embed as a baseline and compare retrieval recall on your own documents.
Can I build an Arabic app on the free plan?
Yes. The free plan includes two free models, plugsky-micro and plugsky-lite, with no card required. Move to a paid plan when you need the higher tiers or more capacity.
How do I evaluate Arabic quality objectively?
Build a scored eval set from real user prompts covering MSA, dialect and code-switching, then compare every candidate on the same inputs with a shared rubric and human review of failures.
Is there a trial for the paid models?
Yes — a 14-day full-access trial lets you evaluate the paid tiers before committing. See the live pricing page for current plan details.
How is pricing structured?
Self-serve plans are flat monthly with fair-use usage rather than per-token billing. Check the live pricing page for current plans and limits.