First of all shout out to Aiden/Antirez & geniuses at the Nvidia community threads. I'm merely claude-vibing off of their works. That a said, i thought i'd share recipes & learnings & benchmarks so far on running big MOE models on two dgx sparks at a reasonable speed for agent use: https://github.com/elsung/dgx-spark-deepseek-v4-flash The kicker here is that you need 2 DGX sparks to really get the speed we need, and you have to spend the $180 on that single cable for 200G...