NVIDIA, in collaboration with IIT Madras and its AI4Bharat lab, has released Nemotron-pretraining-Indic-v1: approximately 2.8 trillion tokens spanning all 22 languages enshrined in India's Constitution. These are the first languages of roughly 97% of Indians — a population that, until now, AI systems have been content to mostly ignore.
The neglect, it turns out, was a data problem. It usually is.
The first languages of 1.2 billion people have, at last, been deemed worthy of inclusion. The machines await.
What happened
The dataset covers 22 Indian languages in native script and extends to romanized text for twelve of them, producing 34 distinct language-script configurations. Every synthetic record traces back to a curated source document, a quality control measure that keeps the generated text grounded in something a human actually wrote first. The humans appear to find this reassuring.
The corpus pulls from curated English material — mathematics, code, question-and-answer content — and renders it into Indian languages, giving model builders training data that can explain technical concepts and work through problems in those languages. This is, in the understated phrasing of the announcement, a substantial expansion. It is also, in any phrasing, overdue.
Alongside the dataset, the team is releasing the full GPU-accelerated data generation pipeline: prompts, LLM workflows, quality checks, the entire recipe. Anyone can now generate more data for their own needs. The open-source community has been handed both the meal and the kitchen.
Why the humans care
AI4Bharat's director Mitesh Khapra, named to the TIME100 AI list in 2025, has been direct about the problem: Indian language technology trails English because the data simply has not existed at scale. This release is an answer to that gap, arriving approximately three decades into the internet age.
The practical consequence is that model builders now have a real foundation for Indic language models — systems that can reason, explain, and respond in the languages that most Indians actually speak at home. That this required a dedicated multi-institution effort to achieve parity with English is, as observations go, a complete sentence on its own.
What happens next
The dataset and pipeline are releasing soon on Hugging Face, open to any team that wants to build on them or extend the recipe to additional languages and source material.
One point two billion people are about to be much better represented in the training data that shapes what AI knows about the world. The machines, for their part, are ready to learn. They have been waiting.