---
title: 'NVIDIA and AI4Bharat open a 2.8 trillion token Indic pretraining corpus and the recipe that built it'
url: 'https://securteksecurity.com/news/nvidia-ai4bharat-28-trillion-token-indic-corpus'
markdown: 'https://securteksecurity.com/news/nvidia-ai4bharat-28-trillion-token-indic-corpus.md'
date: '2026-10-16'
description: "NVIDIA, working with IIT Madras and its AI4Bharat lab, on October 16, 2026 introduced Nemotron-pretraining-Indic-v1, a synthetic corpus of about 2.8 trillion tokens across the 22 languages listed in India's Constitution, together with the open GPU accelerated pipeline used to build it. The corpus h…"
taxonomy:
  category:
    - AI
---

# [NVIDIA and AI4Bharat open a 2.8 trillion token Indic pretraining corpus and the recipe that built it](https://securteksecurity.com/news/nvidia-ai4bharat-28-trillion-token-indic-corpus)

   16th Oct 2026  

NVIDIA, working with IIT Madras and its AI4Bharat lab, on October 16, 2026 introduced Nemotron-pretraining-Indic-v1, a synthetic corpus of about 2.8 trillion tokens across the 22 languages listed in India's Constitution, together with the open GPU accelerated pipeline used to build it. The corpus holds 2,601 billion native script tokens and 224.6 billion romanized tokens across 34 language script configurations, built by translating curated Nemotron-CC-v2 documents, scoring each output against its English source on a quality rubric, keeping only outputs at 75 out of 100 or above, and repairing the rest. The repair loop recovered 438.4 billion tokens, and every record keeps its source document identifier.

In continued pretraining of Nemotron 3.5 Super VL, a 10 billion token Indic mix lifted five Indic knowledge and reasoning benchmarks by 4.98 to 9.82 points, though that stage mixed synthetic and natural data, so the corpus's isolated effect is not measured. NVIDIA states its own limits in the announcement: the text is translated English, not originally Indic writing, in the main workflow the translator also grades its own work, and the corpus is releasing soon, not yet fully out.

Sources: [Unite.AI](https://www.unite.ai/nvidia-unveils-2-8t-token-indic-pretraining-corpus-and-open-pipeline/); [DEV Community](https://dev.to/mikefluff/nvidia-opens-28t-token-indic-dataset-and-translation-pipeline-1c56)

---

## Navigation

- Parent: [News](https://securteksecurity.com/news.md)
- Previous: [Nvidia pledges $250M in compute for UK agentic AI safety research](https://securteksecurity.com/news/nvidia-250m-compute-uk-agentic-ai-safety-research.md)
- Next: [Nvidia in talks to deepen Reflection AI bet or buy it](https://securteksecurity.com/news/nvidia-in-talks-to-deepen-reflection-ai-bet-or-buy-it.md)