Small Gemma 4 tokenizers were recently updated by Google - fix for an issue that caused nonsense output when encountering certain tokens
Gemma 4 E4B as an example, but the same fix was applied to E2B and 12B too. Not sure if this will trickle down to GGUFs