Local Models, Measured

Two DGX Spark nodes in the home lab, and the question of what they are actually good for once you stop reading benchmarks and run your own. Posts here are decisions I made with the numbers that made them: model selection gates, a fix lane that lost to the cheapest hosted model, quantized MoE serving, and the drift that would not reboot away.

Last updated:

  1. The Cheapest Hosted Model Beat My GPU
  2. What a Gate Report Is For Coming soon
  3. A 4-Bit MoE With a Draft Head Wins on One Spark Coming soon
  4. Same Weights, Different Answers Coming soon
  5. Two Sparks, One 284B Model Coming soon