NVIDIA and AI4Bharat outline AI data for 22 Indian languages

NVIDIA and IIT Madras’ AI4Bharat lab have outlined a large training-data resource covering 22 Indian languages. Their October 16 announcement describes roughly 2.8 trillion tokens of text, alongside a reusable process for generating and checking translations. The update says the full dataset is coming soon.

Called Nemotron-pretraining-Indic-v1, the corpus includes text in native scripts for all 22 languages and romanised versions for 12. Romanised text represents an Indian language using Latin characters; it is not the same thing as translating that language into English.

The material is generated from selected English documents through translation and automated quality checks. NVIDIA says the process retains source identifiers so outputs can be traced back. Its announcement also acknowledges that translated data cannot replace material originally written in Indian languages or guarantee coverage of local knowledge.

A separate NVIDIA technical update explains how Indic data has already been used in Nemotron 3.5 Super VL. A continued-training stage added 10 billion Indic tokens drawn from a 12-language subset and naturally authored text. That training amount is separate from the much larger inventory described for the full corpus.

The company reports improved results on several Indic evaluations, but those are its published test results, not a promise that every answer in every language will be correct. For developers, the practical interest is access to more training material and a documented method. For everyday users, this announcement is not a new consumer app or an immediate feature rollout.