The core model seems to have a lot learned behaviour that it applies to all (latin) languages. One is that people sometimes fail to hit space between words and when that happens, in most cases, people actually wanted to insert a space between the two distinct words, and this behaviour has carried over to other languages.

Also In Swedish splitting the words is many cases not incorrect, but just changes the meaning.

For example is Swedish an "English teacher" (Engelsk lärare) would be a teacher that is from England, while an "Englishteacher" is a teacher the teaches the subject English.

So "He is an English teacher" and "He is an Englishteacher" would both be valid sentences in Swedish, but the predictive text model seems to assume you wanted the first one.