[Qwen3.6 35b a3b] Used the top config for my setup 8gb vram and 32gb ram, and found that somehow the Q4_K_XL model from Unsloth runs just slightly faster and used less tokens for output compared to Q4_K_M despite more memory usage
Config CtxSize: 131,072 GpuLayers: 99 CpuMoeLayers: 38 Threads: 16 BatchSize/UBatchSize: 4096/4096 CacheType K/V: q8_0 Tool Context: file mode (tools.kilocode.official.md) Metric M Model XL Model Difference Avg Tokens/sec 28.92 29.78 +0.86 (+3.0…