A model weight quantization algorithm reported in late March 2026 to have approached the information- theoretic compression ceiling for large language model inference. If confirmed, this marks a boundary condition in the memory-efficiency arms race rather than a new engineering record.
