It is absolutely true (I use these models on other languages regularly) and is also clearly explained elsewhere here as an artifact of the tokenizer optimizing

It is absolutely true (I use these models on other languages regularly) and is also clearly explained elsewhere here as an artifact of the tokenizer optimizing for the English heavy training corpus.