MambaByte ⁽¹⁾

Token-free LMs enable faster and fairer NLP

Token-free language models that operate on raw bytes instead of subword tokens can remove tokenization bias, improve computational efficiency, and enable faster text generation.