How about context. We have these not-so-new gadgets made by design to predict the next word, I mean those LLMs... a local tiny model should be able to beat those dumb GBoard predictions any time (and a note for Google: if GBoard uses already such a local predictor, just throw it away, it's garbage)
From https://swipe.futo.tech/:
The ContextLM model is a very small language model that is trained for a single language. It's used to improve the quality of predictions by eliminating nonsensical words given the preceding words in the sentence. It only requires text data for training.
So it would need one model per language? Not impossible (for me)...