I once runned some models on my phone thruh termux. I tried to run Llama 3.2 with 1 and 3B parameters and run pretty well, i tried 8B and was slow. I tried deepseek-r1, 1.5B and run well, 7B was slow.

For text prediction llama 1B may be enough

Now, this is on a 300/400€ phone (Honor magic 6 lite)