I think Zig being a new and a low-resource language (not enough training data) is also one of the reasons Bun decided to use Rust as large language models will simply not be as good no matter how good the model is.

My experience was that LLMs performed at Zig and Rust at about the same level? That is: worse than at something like JS or C or Python, but well enough to write code that works. And they only got better since I first tried them.

Intuitively, Rust would perform worse because ownership is a different paradigm, while Zig is far more similar to things like C/C++. But I guess that's canceled out by the disparity in the language resource base.