I don't think "compressed" is a good way to think about this because the process is lossy. The model can't be "decompressed."

If you think this way, hashing is a copyright violation.