[littlestymaar]

There are models down to 0.6B and you can even run Qwen3 30B-A3B reasonably fast on CPU only.