I spent the last week adding support to DwarfStar and testing two ways to run it: SSD streaming on one 128 GB machine, or distributing the model across two nodes over TCP or RoCE.
1 Like
Great video as usual! Did I miss a comparison to the Halogen-Flash server?
I only bought one machine and now they are far above my budget.
So how are the tradeoffs to the smaller Qwen3.8-Flash-Next model?
I made a separate video on Halogen, ultimately it’s a different model also in size, Qwen 3.8 Flash is faster, this will be more capable overall, but you’ll get a lot done with Qwen 3.8 Flash Next, it’s probably one of the best models for a single Strix Halo right now.