Qwen3.6-35B-A3B Hits 33 T/s on 6 GB VRAM

Hot take: this reads less like a model benchmark and more like a reminder that local LLM performance is often an engineering problem disguised as a hardware problem.

• –The strongest practical finding is that clean IQ4_NL and APEX I-Compact beat “faster-looking” dynamic quants once CPU offload enters the picture.
• –Ngram speculative decoding appears unusually effective for coding-agent workloads, with the author reporting very high draft acceptance and a peak around 33 tok/s.
• –The automated overnight search didn’t uncover anything better than the manually tuned config, which suggests the last-mile gains were mostly from targeted human debugging.
• –The `--poll 0` result looks suspicious: it may be a measurement artifact or a workload-specific anomaly rather than a general optimization.
• –The battery result is the most surprising part: sustained 10 tok/s on a five-year-old laptop is genuinely useful for portable local inference.