I literally just asked that question recently here and got lots of really thoughtful answers.
What are you hoping to use you DGX spark for? The INT4 AutoRound recipe outlined above certainly works with opencode and rust development. Not everything - I use openrouter models for research and planning commands, but my spark seems to handle maybe 3/4 of the implementation work for me right now.
Your milage may vary.
– Oh I didn’t get the spark for coding - but it seemed like the only way I could ground truth the performance. I don’t want to run my own harness on one of these models and not have a decent idea what I am evaluating against. I am not a huge fan of benchmarks. There is a lot of overfitting. I am not interested in speed for speed’s sake. I am interested in long running concurrent self-hosted workloads that run reliably on a properly evaluated harness.