It wasn't a cheap solution but I added an Asus GX10 (DGX Spark variant) and it's been awesome. Huge 128GB of VRAM lets me run Qwen 3.6 35B-A3B Q5 with Vision mod, ComfyUI, a Tdarr node with 2 possible concurrent transcodes, an Immich ML container, and I've still got roughly half the VRAM leftover for other models to play with. Is it the fastest? Nah, but it's already let me take the Arc GPU's out of my MS-01 Proxmox nodes and throw in a couple of 25Gbps NIC's. The software stack on the GB10 platform is changing weekly and seems to be getting better. I did have to build custom Docker containers for Immich ML and some other things I'm playing with, simply because ARM support isn't universal.
It's been a fun learning experience to see how to build out useful AI at home. Next step is seeing if it can automate some smart home devices.
It's been a fun learning experience to see how to build out useful AI at home. Next step is seeing if it can automate some smart home devices.
