I received the DGX Spark on Friday afternoon. Booted it with a Bluetooth mouse and keyboard and a portable 4k monitor and it started up and booted without issue, taking two boots to finish updating.
I installed ComfyUI using the instructions on the DGX page and ran a few images using SD1.5 model. It runs about 3x faster than my 3070 TI laptop.
I installed Ollama and downloaded gpt-oss:20b and gpt-oss:120b models. The 20b model was impressively fast compared to my laptop, taking 4 seconds to think and another 8 seconds to generate a 500 word response. The gpt-oss:120b model took 4 minutes to load and thought for 7 seconds and another 14 seconds to generate a 500 word output.
I installed Flux Krea and it is about 2.5x faster than my 3070 TI for the same workflow. Installed WAN2.2 I2V GGUF models and it is just under 2x faster than my laptop for a 480x720 96 frame video. I also ran a 720x1080 video which my laptop can’t do as it lacks sufficient memory.
Out of curiousity, I tried to follow the playbook exactly and keep getting an error message saying the hashes don’t match on pytorch. I cleared the cache, turned off the cache, but keep getting the same error message. Did you have to do something for this or did it just work?
I don’t recall any error messages during the installation of either ComfyUI or Ollama, though I seem to recall (and have since forgotten) that there was a typo in either the ComfyUI or Ollama installation that I fixed myself during the installation, but regrettably, can’t remember now.
Well it was installing pytorch with the 12.9 version that is in the instructions. I get a hash mismatch error. I instead installed the cuda 13.0 version of pytorch and that went through fine…. Go figure.
Make sure the dgx spark is updated. I used the dgx dashboard for updates.
I just installed comfyUI without any problems, I installed the gpt-oss:20b model, and it’s running fine. Pretty impressive.
In my case I fully updated after first launch and ran into that exact hash error. I did the recovery procedure multiple times since then and it worked.
There are also dockerized ComfyUI packs out there, for those that might end up in the Python dependency hell while trying to find the right package versions.