Qwen3.5-35B-A3B Hits 9 TPS on MacBook Air

Hot take: this is a useful local-LLM hack post, not a model launch, and the real value is the memory-mapping workflow rather than the raw speed claim.

• –Strong practical signal for people trying to run oversized GGUFs on Apple silicon with limited unified memory.
• –The key takeaway is the `--mmap` setup; the rest is mostly tuning flags and a specific quant choice.
• –The 8.9 TPS number is anecdotal and hardware/model-quant dependent, so it should be read as a benchmark note rather than a repeatable spec.
• –Good fit for builders experimenting with llama.cpp, local inference, and Mac optimization.