In the project that I am involved[0], I have been doing stress tests with very large JSON messages. After evaluate a few options I came up that the project 'JSMN' was good enough for my purposes.
https://github.com/zserge/jsmn
I am not sure if there are Python or another language bindings available but that library is very clean and simple to use. It does not perform memory allocations as you (the caller) have to provide the buffers so it gives you good flexibility.