About this site

What is a glitch token

Glitch tokens are vocabulary entries that behave like faulty input units. Training-data gaps or encoding conflicts can make a model garble, repeat, evade, or invent output.

Every entry lists the raw token string, its publicly documented behaviors, the affected model families, and source links — ready to copy and test yourself.

Behavior types

Under-trained: Repetition or embedding tests indicate that the token received too little model training.
Unspeakable: The model evades, substitutes, or cannot accurately repeat the token.
Polysemantic: The same token is inconsistently interpreted as several unrelated meanings.
Fingerprint: The token can help identify the model family or tokenizer behind an API.

Editorial principles

Only publicly documented cases are included, with source links on every entry.
Recorded behaviors come from public tests at a point in time; current model versions may no longer show them.
This site runs no live detection — to test an API, use the tokens from the catalog yourself.

Sources

Plain-text versions for LLMs are also available: /llms.txt · /llms-full.txt