Lawyer runs DeepSeek-V4.1-Flash on 12 used Nvidia V100 GPUs
Second-hand datacentre GPUs plus aggressive expert quantisation can still beat newer consumer units like the DGX Spark on throughput per dollar, but only at the cost of 2,100W power draw and noise that rules out desk-side use. Builders weighing local inference for large MoE models should compare this V100 route against renting GPU capacity, especially where electricity is costly. The published per-context benchmark table is directly reusable for planning similar PP/TP cluster builds.
IT之家Source