Qwen3.5-9B hits performance wall in llama.cpp

Qwen3.5's "thinking" capabilities and hybrid Gated DeltaNet/MoE architecture are currently outstripping local optimization efforts, causing a performance "cliff" on consumer hardware.

• –Architectural complexity (Gated Delta Networks) requires specific llama.cpp updates that are still maturing, leading to high CPU overhead and low GPU utilization.
• –Default "reasoning" modes add significant token overhead, making the model feel slower than dense 8B-9B counterparts despite superior benchmark scores.
• –High VRAM usage for the MoE layers often triggers silent system memory fallbacks on 16GB cards, slashing speeds by up to 70%.
• –Early optimization fixes include reducing --ubatch-size to match GPU cache and explicitly disabling the reasoning budget for standard chat tasks.