Dante-2B debuts as bilingual Italian-English LLM

Training a 2.1B model from scratch for a specific language pair on consumer hardware demonstrates that architectural focus can outperform generic fine-tuning. The custom tokenizer reduces overhead for Italian text by handling morphology correctly, while from-scratch training avoids "translationese" artifacts. Using specialized corpora like FineWeb-2 IT ensures linguistic nuance often lost in large multilingual models.