Pashto and Dari in AI: What Works in 2026

Which Models Handle Dari and Pashto

Why Qwen? Its multilingual training includes Dari and Pashto — something most open families skip. See the Mistral vs Qwen comparison for why Qwen wins the multilingual argument. For translation pipelines, LLM-based translation with glossaries beats generic MT for formal register.

Translation Reality Check

• Dari ≈ Persian/Farsi — benefits from the large Persian training footprint; formal and technical text translates well.
• Pashto is genuinely low-resource — formal text is workable, but dialectal, idiomatic, and region-specific Pashto still needs native review.
• Best practice in 2026: LLM translation with glossary constraints (terminology locked), human review for anything official, and faithfulness-style checks for consistency.
• Direction matters: Dari/Pashto → English is usually better than English → Dari/Pashto. Test both before promising anything.

RAG in Dari and Pashto

The pipeline is the same as any language — with one critical difference: the embedding model must understand the languages, and many don't. The 30-minute RAG pipeline with BGE-M3 is the right local stack:

• BGE-M3 — dense + sparse + multi-vector, strong multilingual coverage including Dari/Pashto-adjacent languages. It's the embedding workhorse in the ranking.
• Test retrieval, don't assume it: build Dari/Pashto questions and measure recall@k before trusting answers.
• Tokenizer note: Arabic-script text tokenizes inefficiently in some models — a Dari chunk may consume 1.5–2x the tokens of the same English text. Budget context accordingly.
• Generation: Qwen3.5-9B answers in Dari/Pashto natively; pair with grounding prompts for document Q&A.

Speech: STT Works, TTS Is Thin

Building a Local Dari/Pashto Assistant

1. Model: ollama pull qwen3.5:9b (or qwen3:8b for smaller hardware).
2. Embeddings: ollama pull bge-m3 — index your Dari/Pashto documents (see LangChain RAG).
3. Grounding prompt in Dari/Pashto with an English fallback note — the grounding rules apply identically.
4. Glossary for translation: lock domain terms (legal, medical) so they survive translation — this is the glossary technique.
5. Evaluate with native speakers: generic benchmarks barely cover these languages — your own test set is the only honest metric.

Frequently Asked Questions (FAQ)

Which AI models handle Pashto and Dari well in 2026?

Qwen-family models are the standout — Qwen's multilingual training covers Dari and Pashto with usable quality, and Qwen3.5-9B is the practical local pick. For translation specifically, test NLLB and Google's models for Pashto. Frontier cloud models handle both languages competently; local 8–9B models are surprisingly good.

Can I build a Dari/Pashto assistant that runs locally?

Yes — Qwen3.5-9B (or Qwen3-8B) via Ollama answers in Dari and Pashto, and combined with a local RAG pipeline it can answer questions over Dari/Pashto documents. The multilingual-translation post in this blog is the step-by-step build.

How good is machine translation for Pashto in 2026?

Much better than a few years ago, but Pashto remains a low-resource language: formal text translates well, while dialectal, idiomatic, and region-specific Pashto still needs human review. Dari (close to Persian/Farsi) translates more reliably. The best practice: LLM translation with glossary constraints and human review for anything official.

What is the challenge with RAG in Pashto and Dari?

Embedding models must actually understand the languages — many general embeddings are weak in Dari/Pashto. BGE-M3 is the strong local choice for multilingual retrieval; test retrieval quality before trusting it. Chunking follows the same rules as any language, but tokenizers vary in efficiency on Arabic-script text.

How do I evaluate a Dari/Pashto model?

Build a small test set of real tasks in each language (translation, Q&A, summarization) and have native speakers score them — generic benchmarks barely cover these languages. For retrieval, use the recall@k method from the RAG evaluation guide with Dari/Pashto questions.

What about speech for Dari and Pashto?

Whisper-family models support both languages for speech-to-text, with quality that improves at larger sizes. TTS voices for Dari/Pashto are sparse — check Piper's community voices and newer multilingual TTS; expect to test multiple options.

Sources & Further Reading

• Building a Multilingual Translation Pipeline with Local LLMs
• Mistral Small vs Qwen: Local Model Showdown
• Top 10 AI Models Under 12B Parameters
• Top 10 Embedding Models for RAG in 2026
• Voice Assistants You Can Build with Local AI
• Lawyer Assistant: privacy-first legal RAG