Qwen3.6-27B MTP tuning hits 50 t/s

This is a deployment win, not just a model win: the same 27B checkpoint becomes much more usable once MTP survives quantization and the inference stack actually exploits it.

• –The speedup depends on an MTP-aware GGUF; plain GGUF conversions usually give up that headroom.
• –The posted config suggests 100k context is the practical sweet spot on a 3090, not maximum possible context.
• –`--spec-draft-n-max 2` looks like the stable setting here; draft 3 was reportedly too heavy at higher context.
• –`q4_0` KV cache plus large batch and ubatch sizes show this is tuned for throughput, not minimum memory.
• –The broader lesson is that long-context local agents may be better served by "enough" context plus compaction than by brute-forcing huge windows.