Although now that I think about it there is still some element of what's described in this article. There's no visual shape involved in the way I model numbers, but it resonates to think of 7 as "10 with a 3 missing", but also as "5 with a 2 on it". The concepts are built in reference to their closest multiple of 5, and slide between different equivalent forms as necessary in calculations.
By the way, the way I do mental math without images feels like it is using sounds and words for the short-term storage and recall. The language brain seems good at putting something aside for a minute and then bringing it back afterwards with a low chance of error, like repeating something someone just said back to them verbatim even though you weren't really listening.
The one method I am sure _doesn't_ work well for mental math is picturing the grade-school algorithms on an imaginary sheet of paper. For whatever reason it is very error-prone. I once did an informal (definitely unscientific) survey on this (30 or so people IRL plus like 100 reddit users) and iirc there was a strong correlation between "imagining the pen-and-paper algorithm", "being bad at mental math", and "not liking math". Wish I still had the data from that -- all I remember is roughly confirming my hunch that those were related. I also wrote a blog post about this a few years ago (https://alexkritchevsky.com/2019/09/15/mental-math.html) but I wish I had included the survey information in there, it would have been much more interesting.