Unsloth Qwen3.6-35B-A3B GGUF hits 44 t/s

This is the kind of result that matters more than synthetic leaderboard noise: it shows a large MoE model can be made usable on prosumer hardware with careful offload and cache settings.

• –The headline number is impressive, but it depends on a hybrid GPU+CPU setup rather than pure VRAM residency.
• –Q8_0 caches and offloading 26 MoE experts to CPU are doing a lot of the heavy lifting here.
• –The post itself hints that llama.cpp may outperform LM Studio for this workload, so the real story is deployment efficiency, not one fixed speed figure.
• –The 128K context claim is the key practical signal: this is about keeping long-context local workflows alive on midrange hardware.
• –For local AI builders, this reinforces that model choice and runtime tuning can matter as much as raw GPU tier.