[kmori_de]

True. I suspect it's still hard to tell whether the bottleneck is the language itself, the tokenizer, or just the overwhelming amount of English training data.