A developer built a custom CUTLASS K=64 kernel to fix broken SM120 MoE GEMM tiles on 4x RTX PRO 6000 Blackwell, achieving 282 tok/s on Qwen3.5-397B (up from 55 tok/s in WSL2). The progression — WSL2 → native Linux → driver optimization → custom kernel — maps a reproducible path for others hitting the same wall. Signals that Blackwell MoE performance requires kernel-level work beyond stock llama.cpp.