Qwen3.6 27B MTP hits 54 t/s

This is a strong reminder that speculative-style decoding can matter more than raw quantization tweaks when you're trying to make a 27B model feel interactive on older datacenter GPUs.

• –The reported jump from ~30 t/s to ~55 t/s is big enough to change the ergonomics of local coding workflows, not just benchmark bragging rights
• –The setup still degrades to 40-45 t/s past 50k tokens, so long-context behavior remains a real constraint even when the headline speed looks excellent
• –The user saw solid behavior on tool calls, sub-agents, and code review/refactor tasks, which is the right test for a coding model
• –This is still a single-user report, so reproducibility will hinge on the branch maturity, prompt mix, and whether the MTP GGUF is as stable across workloads as it looks here
• –The result also shows V100-era hardware can still be surprisingly competitive for local inference when the stack is tuned well