[SparkyMcUnicorn]

You can use 0.6B for speculative decoding on the larger models. It'll speed up 32B, but slows down 30B-A3B dramatically.