A language model that reads and predicts text as small units called bytes.
Most readers see words. This one is like a LEGO kid who checks every tiny brick, even the goofy emoji brick.
It skips the Tokenizer and handles many languages or messy text. Its input gets longer, so it can run slower.
Tokenizer
It reads bytes directly. It does not need a Tokenizer first.
Token
It uses bytes instead of Tokens. This usually makes sequences longer.
LM
It learns language by predicting the next byte.