I'm sure you could even make T5 though it probably works less well with only 5 keys.
For those who don't remember: T9 would have one key for "abc", own for "def" and so on. You didn't have to tap the first two key twice to type "be". You would simple tap the first and the second key, then it would look for words matching "(a|b|c)(d|e|f)" and conclude the most common such word is "be".
I think you could further argument it with grammar awareness, so really you tapped the length of the word. For ambiguous words you would have a key to switch between words.
Everything was deterministic, so you'd quickly be able to type without looking.
For latin script you can combine them: 1 for a or b or c, swipe left for a, swipe up for b, swipe right for c, etc.
But the upside to prediction is that you don't even need to know how to spell perfectly.
Maybe letters that sound similar could be on the same key... If you were to reinvent the system.
Here is an emulator: https://www.sainsmograf.com/labs/t9-emulator/
Type `422537` to receive `hacker`. You hit `2` twice to get `ac`, not `b` - and it picks `ac` given the context.
I think the only limit is the time it takes to train a human to write using such keyboard. And perhaps more problematic how bored said human would get, and hence, give up :)