Llama 3.2 punches way above its weight. For general "language manipulation" tasks it's good enough - and it can be used on a CPU with acceptable speed.
How many tokens/s?
How many tokens/s?