NVIDIA and AI4Bharat released Nemotron-Pretraining-Indic-v1 on October 16, 2026, saying it provides a 2.8 trillion token dataset for 22 Indian languages. It is the company's first major release of Indic pretraining data since the launch of the AI4Bharat lab in 2025.
NVIDIA reported 2,601.0 billion native-script tokens and 224.6 billion romanized tokens, measured using the Gemma-4 tokenizer. That compares with 1.2 billion citizens in India who speak the 22 languages covered by the dataset.
Nemotron-Pretraining-Indic-v1 is built on a GPU-accelerated data generation pipeline and targets Indic language pretraining. Availability begins with the release of the dataset, initially for model builders and developers.
"This data brings a broad mix of curated English material into Indian languages, giving teams more training material for building models that can explain technical concepts," said Mitesh Khapra, head of AI4Bharat. The dataset also includes native-script data for all 22 languages and romanized text for twelve, spanning 34 language-script configurations.
The announcement follows the launch of the AI4Bhar, a research lab focused on open-source datasets and models for Indian languages. NVIDIA and AI4Bharat said the release represents a substantial expansion of the open synthetic Indic training data that will be available to the community.
NVIDIA did not say how the dataset will be used beyond its initial release, and raised the open question of how to ensure the data remains accessible and usable for future research. The company plans to build on this corpus in future Nemotron models.
Source: huggingface
