I suggest adding flag - `full: bool` by default it is False. And if it is True, then it runs: https://github.com/numpy/numpy/blob/main/doc/neps/nep-0001-npy-format.rst The whole point is that it should be implemented in C. Currently: ``` def save(a): m = io.BytesIO() np.save(m, a) # / np.lib.format.write_array return m a = np.ones((1000, 100)) %timeit save(a) 90.8 µs %timeit a.tobytes() 37 µs ``` So... For extra few bytes performance decreases almost 3 times. Mostly due to copy operations. There are stack questions regarding this. Also, various libraries implementing custom types in messagepack and json serialisers, etc... Such improvement would benefit performance and improve simplicity in various places. Regards, DG