May be worth updating the comment at the end of the README, since it currently states that Serde parsing to a struct is faster--I don't believe that is the case anymore.

It depends.

• canada.json is peculiar because it has a vector of two-element vectors (points). Parsing to structs will win here because it can read the point as two f64 on stack, while vector always has to make a heap allocation.
• citm_catalog.json is slightly less peculiar, but it contains one giant object that serves as an associative array of id => proper object mapping. I haven't looked into how serde is representing it exactly, but since the ids are in order I guess a BTreeMap with u64 keys would be pretty fast.
• twitter.json is just json objects in array, this is most similar to what the log benchmark is testing. That said, the log benchmark doesn't have any difficult things, strings are short and I can avoid heap allocation on most of them, this json has a bunch of long unicode strings, so the gap is smaller.

That said, yeah, rewording might be in order.

test json_rust_stringify                 ... bench:       1,354 ns/iter (+/- 497) = 446 MB/s

😮