“Compressing large language model weights by 16x to 100x can reduce the time for an attacker to illicitly transmit them from a server from months to days.”