When you make an NVFP4 (or most other forms of quantization these days), you choose certain parameters to leave unquantized, or quantize less aggressively, or use certain scaling methods that scale them in a way that quantization error is minimized. You do this by measuring how much those parameters affect next token predictions on a calibration dataset, and adjusting them until you get the best performance possible.

So each of these different quantizations may use different calibration datasets. And the calibration dataset can affect how well the quant works for different tasks; for example, if you use an English-only calibratioin dataset, you may lose more performance on other languages than you lose on English. Or if you only use a QA based dataset, you may lose more agentic performance than a dataset that includes agentic traces with a lot of tool calling.

Then you can test afterwards to see how much the quantization actually affected performance on various task evaluations, or you can compare on metrics like KL-divergence (which are also measured against a particular dataset) from the base model that you quantized from.

This particular NVIDIA quant shows very minimal performance decrease on a number of common evaluations. The others cited (RedHat and Unsloth), don't show the same kind of benchmarks, so it's hard to compare.

Unfortunately, there's aren't very good independent benchmarks comparing different quants of the same models. You kind of have to just go with the best info you've got, your gut feel, or do your own evals to compare different quants.