The Language-by-Language Reality
| Language | AI support | Notes |
|---|---|---|
| Persian (Farsi) | Good | Dedicated models (PersianMind, Maral), strong multilingual coverage, real benchmarks |
| Urdu | Good–moderate | Large speaker base; solid support in major multilingual families |
| Dari | Moderate | Close to Persian but distinct; benefits from Persian resources and cross-lingual transfer |
| Pashto | Weak–moderate | Fewest resources; improving with PashtoCorp (1.25B words) and dedicated tooling (NLPashto) |
The Pipeline That Produces Usable Quality
Raw model output is a starting point, not a deliverable. The pipeline that makes these languages usable has four stages:
- Normalize the script. Arabic-script languages need consistent normalization — diacritics, character variants, and word boundaries — before anything else.
- Lock the glossary. A curated glossary of domain terms (legal, medical, government) is injected into the prompt and enforced. This is the single highest-leverage step for consistency.
- Translate with context. Feed the model the previous translations and the document's domain, not isolated sentences. Context changes quality dramatically.
- Evaluate and refine. Score against a small, real evaluation set in each language, and fine-tune on recurring failure patterns.
The full technical pipeline — including the code and workflow for glossary-locked translation — is documented in Building a Multilingual Translation Pipeline with Local LLMs.
The Failure Modes to Watch For
📚 Terminology drift
The same term translated three different ways in one document. Fix: glossary enforcement.
🌊 Dialect drift
Dari slides into Persian, or Pashto into Urdu, when the model defaults to its strongest language. Fix: explicit target-language constraints and dialect examples.
🔤 Script errors
Inconsistent characters, broken word boundaries, or mixed scripts. Fix: normalization before and after generation.
🪤 Literal idioms
Idioms translated word-for-word into nonsense. Fix: idiom glossaries and human review of critical output.
Local vs Cloud
For these four languages, local deployment isn't a downgrade — it's often the better answer:
- Sensitive content (legal, medical, government) should never leave your control — local keeps it private.
- Consistent terminology is easier to enforce when you control the whole pipeline.
- Offline and unlimited — no per-word fees, no connectivity requirement.
- Cloud engines aren't dramatically better on Pashto and Dari — the gap you close with glossaries and fine-tuning is often larger than the model gap.
For hardware sizing, see How Much RAM Do You Need to Run Local AI?.
Frequently Asked Questions (FAQ)
Which AI models translate Dari, Pashto, Persian, and Urdu well?
Multilingual open-weight models — Qwen is a strong common choice — handle all four with varying quality. Persian has the strongest support; Pashto and Dari are weaker but usable, especially with glossaries and fine-tuning. Urdu is better served than Pashto in most families.
How do I get consistent terminology across languages?
Use a glossary-locked pipeline: a curated glossary of domain terms that the model must use, plus retrieval over previous translations. This is the single highest-leverage technique for legal, medical, and government translation.
Can I translate these languages locally, offline?
Yes. A 7-14B multilingual model on a modern laptop (8-16GB RAM) translates all four languages offline and privately. Quality improves with glossary constraints and fine-tuning on your domain.
What are the common failure modes?
Inconsistent terminology, dialect drift (Dari vs Persian), script normalization errors in Arabic-script languages, and literal translation of idioms. A glossary plus a human review pass on critical output handles most of it.
When should I use cloud translation instead?
For very large, non-sensitive volumes with tolerant quality needs, cloud engines may be faster. For sensitive data, consistent terminology, or offline work, local pipelines are the right choice.
🌍 Need translation infrastructure for these languages?
I build multilingual AI for Dari, Pashto, Persian, and Urdu — glossary-locked pipelines, fine-tuned local models, and private deployment. Contact me — this is exactly the kind of work I specialize in.