Open-source neural amp modeler, seems to do better than the ToneX:
Neural Amp Modeler is cool, no doubt about that, but a lot of the heavy lifting is done by the PyTorch machine learning library. That's where all the "magic" lies. NAM is basically just about setting things up to feed it data, setting how the library should handle that data and storing the model.
This is not meant to diminish the effort of its developer, but just to highlight how much of it some smart people have built as is by making the PyTorch library and others like it. Even I could - with a good bit of effort - make something like NAM despite the only machine learning programming I've done is make Google's Tensorflow be able to determine if an image is a cat or not, which is basically tutorial level stuff.
By comparison I would not know even where to begin for something like what Line6 or Fractal do with component based modeling. I absolutely do not have the math skills for it. I've never studied computer science, my degree is in digital communications but I've worked as a software developer for about 15 years on all kinds of web/mobile based systems whether it's user interfaces, mobile apps, backends to process data, databases or cloud services.
Even if Fractal says that his component based approach gives more accurate results, ML-based amp modeling is good enough for most users, especially if it comes with a significantly lower price tag. The problem with it is that it can't really do anything new, just copy existing things.
I am quite confident for example UA pedals are ML models because the UA Ruby's treble/bass controls only work on the Brilliant channel - just like the real vintage Vox AC30s. A sensible approach on a digital modeler would be to say "hey, turn those to this setting for authentic operation or use them as you prefer" and offer both because "these knobs do nothing unless you have this channel selected" is not really a great user experience.
By comparison a component modeling system like Strymon Iridium can say "hey, if you want to push this Fender Deluxe Reverb based model to more of a tweed sound, just turn up the mids!" so it can be more than the original. Which IMO is the beauty of digital modeling - it can be authentic if you want, but also go beyond the limits of the gear modeled by giving users control over aspects of the sound.