What is binary quantization and when does it work well?

Binary quantization reduces each vector dimension to a single bit, storing only the sign of each float, giving a 32x memory reduction and enabling distance computation via fast bitwise XOR and population count. It works well when embeddings have high dimensionality and well-distributed values, since random projections preserve relative distances in Hamming space. Binary quantization pairs with a rescoring step that re-ranks top candidates using full precision vectors. Modern embedding models from OpenAI and Cohere are explicitly designed to support binary quantization.