This post was written in my own words, but with AI assistance.
I own two DGX Sparks myself, and the lack of NVFP4 has been a real pain in the ass.
The reason the product made sense in the first place was the Blackwell + NVFP4 combo on a local AI machine with a proper NVIDIA software stack around it. Without that, Spark becomes much harder to justify, especially given the bandwidth limitations and the compromises that comes with it.
The DGX Spark was presented like a finished, premium system where NVFP4 was supposed to work out of the box. It was not marketed like an experimental dev kit where buyers should expect to spend months switching backends, testing builds, setting flags, and relying on community or hardcore fan fixes just to make a core feature work properly.
More than six months in, NVFP4 is still not properly delivered on the Spark. Yes, you can get things somewhat running. But there is a big difference between a feature technically existing and a feature being delivered as a mature, stable, and supported experience.
Right now, NVFP4 on Spark is much closer to the first than the second.
The hardware itself is not the main issue. Spark has potential, and in some scenarios it can perform well. But the overall experience does not match what was implied. At this point, it no longer feels like normal early friction. It feels like NVIDIA pushed the story before the software was actually ready.
So the takeaway is simple:
Do not buy DGX Spark assuming NVFP4 is already delivered as a polished, mature, supported feature.
NVIDIA overpromised and underdelivered on DGX Spark.
Rant over and out.
[link] [comments]
