from ml_dtypes import bfloat16
import numpy as np

x = np.arange(4, dtype=bfloat16)
np.save('out.npy', x)
print(np.load('out.npy'))
# [b'\x00\x00' b'\x80\x3F' b'\x00\x40' b'\x40\x40']

This is due to the fact that np.save saves the dtype via the kind code, and we chose np.dtype(bfloat16).kind = 'V' because such codes are not extensible in numpy.

When reconstructing the dtype, this is the result:

np.dtype(np.dtype(bfloat16).kind)
# dtype('V')

I don't think this is something that can be fixed without deeper changes to numpy itself.