Researchers adapted Nemotron-3-Nano-30B-A3B for eight Indian languages by expanding its tokenizer in place. The vocabulary grew from 131,072 to 163,072 entries through continued byte-pair encoding, a method that extends the existing tokenizer rather than replacing it. This allowed the model to handle new languages without modifying its underlying architecture.
The team used a continual pretraining strategy called AdaptMix, which mixes new-language data with existing data to preserve previously learned capabilities. Because the backbone stays intact, the adapted model retains its original structure and can be used with the same deployment infrastructure.
The work demonstrates a practical path for adding languages to large language models without full retraining or architectural redesign. The paper reports success across eight Indian languages, though details of evaluation and performance are not included in the abstract.