What’s the best harness alternative to ZCode for using with GLM Flash 5.3? Asking for a friend…

With exactly 4 Sparks are switchless recipes a viable solution or is the overhead of that kind of unofficial cluster setup eating enough of the performance that you really should pony up for the switch?

What’s a useful thing to do with a standalone single Spark that can’t yet join the cluster (waiting on cables)? Throw MiniMax on it?

Used these graphs to convince my work to secure 2 sparks for special projects at the Dallas Micro Center. Thanks for tracking this @loktar00.

The theme song for DGX Spark inventory is "The Final Countdown"

This is phenomenal for a locally generated music video. My anthem for today.

Thanks to all who responded to my question about which model is currently best for a small team (2-4) sharing a 2x DGX Spark cluster. There was consensus around these two models: 🥇Qwen 3.8 Flash Next (recommended) 🥈DeepSeek V4 Flash (faster, more concurrency) If you're solo, step up to GLM 5.3 Flash. But it's too slow for sharing.

The last time I checked today, my local Micro Center had 25+ DGX Sparks in stock. Now it's not showing any. Did they really go that fast?

A few days ago after the ZCode incident, I asked people for their favorite alternative harness for GLM 5.3 Flash. Here are the results. I've personally tried Pi, OMP, Hermes, and DSH so far with it and they all seem to work well. Looking forward to trying some of the others. Of course now that ZCode has been open sourced I may give it a second chance.

Anyone with 2 Sparks have more than one person using it at the same time? It's fast enough for my single use with GLM 5.3 Flash, but wondering if I shared it with 1-2 other people if it would be too slow? I could run some tests but just want to see if anyone is doing this.

Another demo, this time by Astra (graphics) and YuE2 (music). Growing up I really wished I could program well enough to make stuff like this (but I was still a n00b).