Back to Project
N

Nemotron Indic Pretraining

Project Oct 16, 2026

NVIDIA and AI4Bharat's announced synthetic Indic pretraining corpus, paired with a public NeMo Data Designer translation and evaluation plugin.

Visit primary link 3 sources attached
NVIDIA's diagram of the twelve-language Indic workflow: translate English documents, score and repair them, then optionally romanize accepted text.
NVIDIA's author-published diagram for the twelve-language translation and romanization workflow. Quality scores are rubric judgments, not accuracy percentages; AITrending did not reproduce the pipeline. Source: NVIDIA and AI4Bharat
self-hostedverification

What happened lately

Nemotron Indic Pretraining timeline

NVIDIA and AI4Bharat at IIT Madras describe approximately 2.8 trillion synthetic pretraining tokens across 22 Indian languages. Their October 16 announcement says the corpus is releasing soon; this record does not establish that the complete corpus is already downloadable.

The public translation plugin runs inside NeMo Data Designer and translates, evaluates and optionally rewrites source documents. Its repository package specifies Python 3.10 or newer, Data Designer 0.9.4 or newer and Apache-2.0 licensing. Model inference and corpus licensing are separate from that plugin license. The announced corpus includes native-script text for 22 languages and romanized subsets for twelve, derived from curated English source material.