LittleBit and LittleBit-2 are NeurIPS 2025 and ICML 2026 methods for compressing large language models to sub-1-bit regimes through latent factorization, binarization, and learned scales. LittleBit-2 improves initialization via Joint-ITQ alignment without adding inference overhead.
CRAM, a new Linux memory compression technology developed by Gregory Price's team at Meta, keeps compressed data in RAM rather than using swap, achieving up to 452x performance improvement over ZRAM. Unlike traditional swap-based solutions like zswap and ZRAM, CRAM uses a private NUMA node to enable faster memory access while maintaining Linux's memory management semantics, though challenges around managing variable data compressibility remain unsolved.