Qwen3.6 35B hits 564/41 on 5070 Ti

The setup leans on llama.cpp flags such as n-cpu-moe, cache-type-k q4_0, and cache-type-v q4_0 to make a 35B MoE model fit. The memory profile is the constraint: 10.8/16GB VRAM plus shared RAM and normal RAM pressure, so this is viable only on a relatively loaded but carefully managed workstation. A 262K context window is impressive, but it makes performance claims highly configuration-dependent rather than broadly transferable. The TurboQuants miss is a useful warning sign that local LLM tuning still has rough edges even when the base model runs well.