Google updated the tokenizers for small Gemma 4 models (E2B, E4B, 12B) to fix an issue causing nonsensical output when encountering certain tokens. It remains unclear if these fixes will propagate to GGUF distributions.
Read original
reddit/r/LocalLLaMA