MTP speeds coding, drags creative output

Speculative decoding is not a universal speed upgrade; it is a workload-sensitive optimization that only pays when the model can reliably predict the next tokens.

• –Coding is the sweet spot: draft acceptance stayed high enough that MTP delivered large gains across every quant tested.
• –Creative writing is the failure case on cheaper quants, where lower acceptance rates make the extra draft pass more expensive than the tokens it saves.
• –Memory bandwidth still matters, but it is secondary to task predictability here; F16 and Q8_0 win big because they are more bandwidth-bound and can better amortize the draft overhead.
• –The finding that `num_speculative_tokens=3` is usually optimal is practical, not just academic: more drafts hurt once acceptance starts sliding.
• –The takeaway for local inference is simple: benchmark on your real workload, not on a generic chat loop, because task mix changes the conclusion more than quant size does.