The user operates a four‑board BC‑250 mining‑hardware cluster linked via 2.5 GbE, using Llama with Vulkan and RPC to reach ~70 tokens/s with Next Flash IQ3_XXS (100k context) or ~145 tokens/s with two Qwen 3.6 35B A3B IQ4 instances (256k context) for under $500. This improves upon earlier results of ~30 tokens/s on Flash Next IQ2 and ~80 tokens/s on a single Qwen 3.6 35B model. All boards are housed in their original ASRock 4U12G BC‑250 chassis.