Related work
Lossless compression of model weights is not new. What Glyd adds is the combination: the smallest size, the KV cache too, and the same result every run.
The work Glyd builds on
- DFloat11 (Zhang et al., NeurIPS 2025): Lossless bf16 weights, about 30% smaller, outputs bit for bit; its authors report about twice bf16’s time at one sequence.
- ZipServ (ASPLOS 2026): Lossless bf16 weights for serving: up to 30% smaller, up to 2.21 times cuBLAS’s speed per product and 1.22 times vLLM’s end to end on GDDR GPUs.
- Approaching Shannon Bound with Lossless LLM Weight Compression (Tan et al., ISCA 2026): Lossless weights for bf16, FP8 and integer formats within 0.01 to 0.05 bits of the entropy bound, in SGLang.
- SplitZip (2026): Lossless compression of the KV cache, for moving it between servers.
- ZipNN (2024): Lossless compression of model weights for storage.
- NeuZip (2024): Compressed weights to save memory in training.
- Huff-LLM (2025): Lossless weight compression decoded in hardware.
- Lossless compression of weights, checkpoints and K/V caches in low-precision formats (2025): Lossless compression of weights, checkpoints and the KV cache in low-precision formats.
- Unweight (Cloudflare, 2026): Lossless compression of MLP weights.
- dietgpu (Jeff Johnson, Meta, 2021): A GPU entropy coder and a lossless codec for float16 and bfloat16, used to compress what GPUs send each other in training.
- fpzip (Lindstrom and Isenburg, LLNL, 2006): Lossless and lossy compression of floating-point arrays by prediction and entropy coding.
- FPC, MPC and others (Martin Burtscher’s lab, Texas State): Fast lossless compressors for floating-point data on CPUs and GPUs, from 2009 on.
Where Glyd differs
- The smallest size. 10.80 bits a weight (32.5% off bf16), where DFloat11 and ZipServ take about 30% off. On the 2026 models measured in vLLM the whole model is 21.3 to 28.2% smaller so far, parts of them not compressed yet (the benchmarks).
- The KV cache too. The cache is held compressed: 31% smaller.
- The same result every run. Run it twice and the output is the same.
- Beyond the GPU. Models on disk, a fine-tune against its base and training checkpoints with their optimizer state are smaller too.
How Glyd works, in short: how it works. What runs on the GPU: the GPU docs.
Side by side
On one 16 GB RTX 4080 SUPER with Qwen3-8B, whose 16.38 GB of bf16 does not fit it:
| Against | Measured | Glyd |
|---|---|---|
| DFloat11, at the same size | Tokens a second, 1 to 64 sequences | 2.5 to 3.4 times DFloat11's |
| ZipServ, product against product | Microseconds a call, 4 matrices at 1 to 64 tokens | The fastest in 10 of 16, in 5% fewer bits |
The numbers, the method and the raw logs: the head-to-head report and the benchmarks.