Try translating Kashmiri with Google Translate and you will find it is not on the list. Google now covers roughly 240 languages, including several Indian ones added in 2024, and Kashmiri is still not among them.
Who does translate Kashmiri
- Microsoft Translator lists Kashmiri (
ks) for text translation. - IndicTrans2 (AI4Bharat) and NLLB-200 (Meta) are open models that include Kashmiri in Perso-Arabic script.
- Bhashini, the Government of India's language platform, offers translation for Indian languages; check its current list for Kashmiri.
- Our translator, NLLB-200 fine-tuned on 93,064 Kashmiri–English pairs, is free to try for research use.
Support is not the same as quality. For Kashmiri, even the systems that cover it make frequent mistakes.
Why quality is low
Translation systems learn from parallel text. English, Spanish or Hindi have billions of sentences available; Kashmiri has a few hundred thousand public parallel pairs at most, spread across a few collections. Kashmiri is also written in two scripts, Perso-Arabic Nastaliq and Devanagari, with spelling that varies from writer to writer, which makes the little data there is harder to use.
What about general AI models?
Large multilingual models see very little genuine Kashmiri text during training. Asked to translate it, they tend to produce:
- fluent English that misses the key meaning;
- transliteration instead of translation;
- hallucinations: confident English with no relation to the source;
- failures on idioms and culturally specific phrases.
In the ICRTC-2026 paper, a dedicated 600-million-parameter translation model beat a fine-tuned 7-billion-parameter general-purpose LLM on Kashmiri, as reported; follow-up work in 2026 is re-testing that result.
What helps
- More and cleaner parallel data, with its sources and licences respected.
- Fine-tuning dedicated translation models rather than relying on general chat models.
- Honest benchmarks: test sentences that no training corpus contains, so scores are not inflated.
- Native speakers checking the output, because automatic scores miss many errors that matter.