NVIDIA researchers together with AI4Bharat at IIT Madras published a Hugging Face Blog post on October 16, 2026 describing Nemotron-pretraining-Indic-v1, a planned open pretraining corpus of roughly 2.8 trillion tokens spanning the 22 languages listed in India's Constitution. The release pairs the corpus with an open, GPU-accelerated data generation pipeline built on NeMo Data Designer so other teams can reproduce or adapt the recipe.
The corpus mixes native-script data for all 22 languages with romanized text for twelve of them, yielding 34 language-script configurations. Authors report about 2,601 billion native-script tokens and 224.6 billion romanized tokens when counted with a Gemma-4 tokenizer. Source material starts from curated English Nemotron-CC-v2 subsets covering web text, question-answer content, mathematics, and code, then moves through document-level translation, quality scoring, and targeted repair.
For twelve languages the workflow uses Gemma-4-31B-it for translation and MQM-inspired scoring, keeping outputs at or above 75 out of 100 and recovering lower-scoring items via feedback-guided repair. Romanization is applied as a second step from accepted native-script text rather than generated directly from English. For the remaining ten languages the team adapts Indic Translate from Bodhan AI with document windowing and script remapping where needed.
The post links the corpus to continued pretraining on Nemotron 3.5 Super VL, where a configured 100 billion token stage allocated 10 percent to a mix of synthetic and natural Indic text and reported gains on MILU, Indic MMLU-Pro, Indic ARC, and Indic GSM8K. Those gains are stage-level and attributed to the combined mixture, not an isolated ablation of the full 2.8T pool.
Openness is the editorial core. NVIDIA and AI4Bharat say they will publish the corpus configurations, dataset card and license guidance, streaming load examples, and the Nemotron Bharat translation plugin so builders can inspect samples, pin revisions, and regenerate domain-specific data. Companion materials include a technical report on synthetic Indic pretraining and a separate Indic-LLM-Eval-Suite effort for benchmarks.
Limitations are stated plainly: translated synthetic text does not replace naturally authored Indic writing, automated judges can miss errors, and source-ID dedupe is not semantic deduplication. Teams are told to combine the corpus with natural Indic text and to track English capability retention alongside per-language quality.
This story is distinct from the earlier Nemotron 3.5 Super VL model announcement already in published news. Here the central event is the open Indic data and generation pipeline, not a new model checkpoint drop.