A new technique called Speculative KV coding enables the lossless compression of AI model KV caches by up to four times. By leveraging a cheaper predictor model to drive an arithmetic coder, the method significantly reduces memory overhead for large language models without sacrificing data integrity.