This is not a new model, and the corpus cannot be downloaded yet
NVIDIA and AI4Bharat at IIT Madras posted on Hugging Face. The structured publish time is 2026-10-16 01:44 UTC, 9:44 a.m. in Beijing. They introduce Nemotron-pretraining-Indic-v1: about 2.8 trillion tokens across the 22 languages listed in India's Constitution, releasing soon. They call those the first languages of 97 percent of Indians, nearly 1.2 billion people at the 2011 Census. That population line is theirs.
The same post says continued pretraining for Nemotron 3.5 Super VL added 10 billion Indic tokens, from a twelve-language subset of this corpus plus natural Indic text. That is a training stage of a model already announced, not another model launch. They plan to build on the corpus in later Nemotron models. A plan is not a training run that has already happened.
[1]
2.8 trillion is an inventory, not a training budget
Figure 1 says the corpus contains 2,601.0 billion native-script tokens and 224.6 billion romanized tokens. Counts use the Gemma-4 tokenizer. A record can be a translation or another script of a shared English source. These are not counts of unique English documents.
Native-script here means the writing system. The text is still translated from English. Romanized means the Indic language written in Latin script, not English. Twelve languages have both, including Hindi, Tamil, Telugu, Bengali, Marathi, and Urdu. Ten others are native-script only. That is 34 language-script configurations. The token counts describe what was kept, including accepted rewrites, not how many tokens a training run spent. A different tokenizer changes the inventory.
[1]On twelve languages, the translator is also the judge
For those twelve, Gemma-4-31B-it translates full documents and then scores each output against the English source. They keep only scores of at least 75 out of 100. That is a rubric score, not an accuracy percentage. Lower-scoring outputs are repaired from the feedback and scored again against the same line. This recovered 438.4 billion tokens, already included in the total. The other ten languages use Indic Translate from Bodhan AI, then Gemma scores them. Those paths produce native-script text only, with no feedback rewrite. Longer Santali documents are translated in overlapping windows and reconstructed. Manipuri is translated into Bengali script, then converted to Meitei Mayek.
Once the release is available, they say to pin a dataset revision. That sentence is still about afterward. Quality checks are mostly automatic, and sampled human review covers languages unevenly. On the twelve-language path, the translator is the judge. Source-ID checks remove processing duplicates. They are not semantic deduplication. Translation widens access to English source material. It does not replace text originally written in these languages, and it does not guarantee local knowledge.
[1]Five benchmarks rose, and the synthetic corpus was not isolated
The technical report compares the checkpoint just before Indic continued pretraining with the checkpoint after that stage. The configured budget was 100 billion training tokens. Ten percent, 10 billion tokens, went to a mixture of synthetic and natural Indic text. The other 90 percent kept the preceding mixture's relative proportions. MILU, Indic MMLU-Pro, Indic ARC-Challenge, Indic ARC-Easy, and Indic GSM8K all rose, by 4.98 to 9.82 percentage points. The comparison is the whole mixed stage, not the synthetic corpus alone. These are stage checkpoints, not two final releases. Indic GSM8K excludes Odia because evaluation exceeded the serving context limit.
Against the previous Nemotron 3 Super, the released Nemotron 3.5 Super VL is higher on five of the six task benchmarks in the figure. IFEval 1.5 dipped 2.04 points. All four IndiVibe dimensions improved. Those are preliminary conditional preference scores, not head-to-head win rates. The scores reflect the full training pipeline, including this continued-pretraining stage and post-training. They are still not an experiment that holds out the 2.8 trillion tokens by themselves.
[1]