Qwen3 NextN MTP guide for llama.cpp

This is the kind of guide power users want and most people will skip: useful, specific, and heavily dependent on a non-upstream branch plus model-file hygiene.

• –The real trick is not just enabling speculative decode, it is keeping the `nextn` tensors intact through quantization so the head still works
• –The performance story is hardware- and tuning-sensitive, with the post’s claims tied to a 3090 Ti, power/clock locks, and q8 KV cache settings
• –It is most relevant to people already running local inference stacks and willing to recompile, re-quantize, and troubleshoot edge cases
• –The guide also signals where llama.cpp is heading: MTP support looks promising, but the ergonomic path is still fork-first rather than turnkey upstream
• –For broader users, this reads more like an advanced operator note than a mainstream release announcement