PSA: For those of you that can't keep up with this forum

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Just a PSA for the community members who, like me, are having a hard time following the posts here.

Nvidia’s forum is built on Discourse, and its MCP works. I just plugged it into grok-build and it works quite well. Task your agent to set it up for you.

This forum remains an excellent source of information and the MCP is a godsend to separate signal from noise.

Here’s a nice sample – I asked Grok to tell me what the community has been saying are good >100B multimodal models to run in my 2x Spark cluster.

image
image2352×999 161 KB

3

Our recently posted deepseek dspark recipe works pretty well on 2x and has been good for me: DeepSeek-V4-Flash-DSpark on 2× DGX/ASUS Spark (GB10 / SM120)

Others on this form have posted different methods of running it as well with varying results. Ours adds a tweak that we found a bit faster on top of stock eugr recipe.

4

Great tip!

For two sparks I’ve had the best experience so far with DeepSeek v4 Flash. This recipe is what I’m running most of the time:

Just about to try the recently released Laguna S 2.1 to see if it can best it but I am skeptical: Laguna-S 2.1 FP8 — Dual DGX Spark (TP2, W4A16 DFlash) — Recipes · How To Spark