FWIW, Qwen 3.8 27B is actually slower with MTP enabled, unless you generate a lot of code with it (and even then it's not obvious, because you'll likely spend m

FWIW, Qwen 3.8 27B is actually slower with MTP enabled, unless you generate a lot of code with it (and even then it's not obvious, because you'll likely spend most time thinking).