Type or paste Thai. Every tokenizer below splits it live in your browser - the same splits the real tokenizers make, checked id for id on 2,114 test texts. Blue rows are this project's retrofits: the same vocabulary size, far fewer Thai tokens.
One box per token. Pick any two tokenizers; click a row on the board to put it on the left.
Tokens are what a model is billed, limited and timed by. Change the assumptions; everything recomputes from the text above.
Context: how many copies of this text fit in an 8,192-token window. Hosted GPT rows are for reference - their tokenizer cannot be changed by users, which is why this project targets self-hosted open-weight models.