I don't really have a favoured solution, as of now, other than not using predictive typing.

I can see that Minuum works well using Bayes to predict a certain english word given the user input.

If you guys generate/use a language model for other languages that would give us around the same accuracy/prior probability as the english model, then I think it shouldn't be a problem.

Having quick settings and options that you can setup and choose which languages to include in the prior probability.

Being able to quickly turn off/turn a language, add/remove languages to the priori, etc.