A LocalLLaMA community post says Google updated Gemma 4 tokenizers to address nonsense output on certain tokens, with the fix applied to E4B, E2B and 12B models.
The LocalLLaMA post gives Gemma 4 E4B as an example and says the same tokenizer fix was applied to E2B and 12B. It does not identify the affected tokens or provide further details about the issue.
The post says it is unclear whether the update will be incorporated into GGUF versions.