Glyd is a lossless compression library and CLI written in Rust with a C ABI, positioned as a drop-in alternative to LZ4, Snappy, and zstd for workloads where storage size and read CPU decide the bill; it adds a record mode (-r) that turns logs, CSV, and JSON lines into typed columns before compressing, and a base mode (--base) that compresses a new file version as a delta against an older one. Its central feature, matching the repository's tagline "Lossless AI compression: 33% less GPU memory, bit for bit," is a GPU component (gpu/) that holds a model's bf16 weights and KV cache compressed in GPU memory and decodes them bit for bit inside the matrix-multiply and attention kernels instead of unpacking to bf16 first. On Lambda Cloud 4x RTX A6000 GPUs (2026-09-26), Qwen3-32B's 65.5 GB of bf16 weights compress to 44.5 GB, fitting on one 48 GB GPU instead of two; on an H100 SXM the same day, Llama 3.3 70B Instruct shrinks from 136.9 GB to 91.9 GB and Qwen3 30B-A3B (MoE) from 59.8 GB to 40.2 GB, both around 32-33% smaller, with MMLU accuracy and perplexity within a fraction of a percent of bf16; a Qwen2.5-7B KV cache at 16K tokens drops from 947 MB to 651 MB. Generation throughput on the same A6000 runs is 1.05-1.42x bf16's tokens per second at 1-32 concurrent sequences, but drops to 0.82-0.86x (slower than bf16) at 64 sequences. On general-purpose data it also decompresses 3-7x faster than zstd at comparable or better ratios while writing somewhat slower than zstd -3. Licensing is split three ways: the codec, CLI, C ABI, and Python/Go bindings are dual-licensed under the BSD 3-Clause License or, at the user's option, GNU GPL version 2 (the same choice zstd offers); the glyd-store crate and the gpu/ weight-compression code are each separately licensed under the Business Source License 1.1, allowing free personal, educational, research, and other non-commercial production use but requiring a paid commercial license for revenue-generating production use, converting to Apache License 2.0 four years after each version's release. A stated limitation: record-mode reads cost roughly 2-2.7x zstd's CPU to rebuild columns, making zstd -3 cheaper at very high read volumes.