I thought a bit about this - I think you could do a lot faster if you had a 'bulk' method to create lots of Python objects in one hit. I don't know it that's feasible, as I don't know much about the innards of Python, but there's no reason one couldn't keep goofing around with SIMD code and branch-free stuff all the way until you've created objects.
This may be ferociously unmaintainable (i.e. depend on things that Python implementations don't advertise as a part of the API) but I don't know. I doubt you're the only one to have this problem.
I think the time is currently best spent trying to push more of the work back into native code, extending the simple filtering that's in there now to also handle mutation (similar to jq), reducing the amount of back-and-forth with the interpreter.
There was no trick I could find (although treating all strings as bytes instead of using PyUnicode_DecodeUTF8/unicode_decode_utf8 was a fairly large gain on string-heavy documents) to reduce the cost of bringing the document into Python land, so I added a quick hack to push the filtering down into C, since most of my use cases only need a small part of the full document.
[1]: https://github.com/TkTech/pysimdjson [2]: https://github.com/TkTech/pysimdjson/issues/22