Changelog
Every release, newest first. The full notes of each, every change with its numbers: its release on GitHub, and CHANGELOG.md.
v0.32.1 Oct 17, 2026
- Gemma 4 writes faster: 12B at wren 81.2 to 88.9 tokens a second on an RTX 4080 SUPER
- glyd fetch MODEL
- glyd verify downloads first; clear stop where a glyd/ repo does not fit
v0.32.0 Oct 17, 2026
- Gemma 4 12B and 31B at :penguin, Google's bf16 weights bit for bit
- The 12B on a 24 GB card, the 31B on a 48 GB card
- glyd run google/gemma-4-12B-it with no level runs penguin
v0.31.1 Oct 16, 2026
- glyd run google/gemma-4-31B-it:wren (and the 12B) works as written
- glyd serve runs Gemma 4 on Glyd's engine
- Downloads go on over IPv4 when IPv6 fails
v0.31.0 Oct 15, 2026
- Gemma 4 31B and 12B at :wren, Google's QAT weights bit for bit at 4.5 bits per weight
- Drafting with Google's draft models: 31B at 78-80 tokens a second on an RTX 4090
- glyd verify MODEL
v0.30.0 Oct 10, 2026
- glyd run MODEL:swift downloads a ready-made checkpoint and starts from it
- Nemotron 3 Super (120B) at :hummingbird on one GPU: 43.6 tokens a second on an RTX 4090
- glyd serve returns XML tool calls as tool_calls
v0.29.4 Oct 9, 2026
- glyd run MODEL:LEVEL: :penguin lossless, :kestrel about 6.5 bits per weight, new :swift about 5.5
- Qwen3.8-27B with :swift on one A100 80GB: 17.4 GiB of weights, 47.5 tokens a second
v0.29.3 Oct 9, 2026
- glyd run --lossy at about 6.5 bits per weight, closer to bf16
- Qwen3.8-27B at 31.3 tokens a second on one L40 and 45.0 on one A100
v0.29.2 Oct 8, 2026
- glyd run --lossy: Qwen3.5, 3.6 and 3.8 at about 6.8 bits per weight on Glyd's own engine
- Qwen3.8-27B in 21.2 GiB of weights at the FP8 checkpoint's speed, one user
v0.29.1 Oct 6, 2026
- More 2026 model families in glyd run, glyd serve and glyd doctor
- More loaded weights compressed with their original values preserved
- Correctness fixes and smaller saved models
v0.29.0 Oct 5, 2026
- Qwen3.5 measured: Qwen3.5-9B's weights 28.2% under bf16's on an RTX 4090 and 21.3% on an RTX 5090
- glyd run and glyd serve start vLLM compiled; Glyd's own engine ships, off unless asked for
- All of Glyd under the Business Source License 1.1; model files smaller with glyd --max
v0.28.0 Oct 4, 2026
- The whole model smaller under vLLM: the embedding and the output layer packed too, Qwen3-8B 32.0% under bf16's weights on an L4
- The smallest layout faster at 1 to 16 tokens a step; saves in format glyd-v4
v0.27.0 Oct 3, 2026
- The lossless KV cache in vLLM: 1.25 to 1.30 times vLLM's tokens in the same memory, bit for bit
- The GPU parts ship compiled, as the glyd-gpu package: the same install and license
- Fixes for transformers 5.18's embeddings and the glyd run window with the KV cache; vLLM serving rows corrected after a benchmark fix
v0.26.0 Oct 1, 2026
- One command to a chat with a model: install.sh, then glyd run MODEL
- vLLM serves Glyd: vllm serve MODEL --quantization glyd
- Faster long prompts on an A100 SXM, a GH200 and an H100 SXM
v0.25.1 Sep 29, 2026
- Long prompts on the L4 and L40S within 5% of bf16's at 8,192 tokens
- Exact mode and the longest prompts faster again on Hopper, measured on a GH200
v0.25.0 Sep 29, 2026
- glyd pack and glyd verify in Rust, with no Python or GPU
- Models saved in the 12-bit layout, loaded as saved
- A faster 12-bit layout on the H100 and A100; pack 12-bit models again
v0.24.0 Sep 28, 2026
- generate() compiled by default on PyTorch 2.13 or later, with no code change
- Faster Hopper prompts; the CUDA 12 library as fast on Hopper as the CUDA 13 one
- Faster long prompts on an A10
v0.23.0 Sep 28, 2026
- The GPU kernels as a CUDA library of their own, for engines in C, C++ and Rust
- Prompts faster on RTX 40-series GPUs, with the same bits
- Every release installed and checked on fresh machines, the ways users install it
v0.22.0 Sep 28, 2026
- Mixtures of experts saved and loaded packed
- Every mixture-of-experts family packs its experts, Llama 4's included
- Less GPU time on H100 and A100 GPUs, and native code for Blackwell
v0.21.0 Sep 27, 2026
- pip install "glyd[gpu]": the GPU kernels in the Python package, nothing to compile
- glyd.from_pretrained() packs a bf16 model on the GPU as it loads, mixtures of experts included
- exact=True for bf16's logits bit for bit, and a model saved packed
v0.20.0 Sep 27, 2026
- The Python package on PyPI: pip install glyd
- Nine more open models measured, nineteen in all
- Faster steps on Hopper from the 12-bit layout
v0.19.0 Sep 26, 2026
- A second GPU layout, 12 bits a weight
v0.18.0 Sep 26, 2026
- The KV cache held compressed in GPU memory, bit for bit
- Larger models in fewer GPUs: Qwen3-32B on one 48 GB GPU, Qwen2.5-72B-Instruct on three
v0.17.0 Sep 26, 2026
- The smallest GPU layout: 10.80 bits a weight, 32.5% under bf16, bit for bit
- The GPU code under the Business Source License 1.1 from this release
v0.16.0 Sep 25, 2026
- Several tokens at once on the GPU, from the packed weights
v0.15.0 Sep 25, 2026
- A model's bf16 weights held compressed in GPU memory, bit for bit
- PyTorch checkpoints compressed
v0.14.9 Sep 25, 2026
- The store at a terabyte: 3.52× fewer bytes than zstd -3
- Model weights and checkpoints: a checkpoint stored against the one before
- The same bytes on every machine
v0.14.8 Sep 24, 2026
- The store at a terabyte: 3.46× fewer bytes than zstd -3
- Record mode writes 1.6–1.9× faster, the same bytes
- Parquet files with snappy or zstd pages opened
v0.14.7 Sep 24, 2026
- The store's put 1.6–4.0× faster, its get 1.1–1.5×
- Faster reads from the command line
v0.14.6 Sep 24, 2026
- Smaller output on binaries with sections of different content
v0.14.5 Sep 24, 2026
- Faster writes at --max
v0.14.4 Sep 23, 2026
- --max writes faster; repeats far apart in a file are found with --long
v0.14.3 Sep 23, 2026
- Fewer bytes on records, at the same speed
v0.14.2 Sep 23, 2026
- Smaller table dumps at the max level
- Every block checked for corruption
v0.14.1 Sep 23, 2026
- JPEG recoding on every core
v0.14.0 Sep 23, 2026
- JPEG files recoded smaller, restored bit for bit
v0.13.4 Sep 23, 2026
- Deflate streams opened and restored bit for bit by Glyd itself
v0.13.3 Sep 22, 2026
- Containers open and close on every core
- An opened container's content read on its own
v0.13.2 Sep 22, 2026
- The default, fast and turbo levels leave containers closed, for speed
v0.13.1 Sep 22, 2026
- Fixes: every file v0.12.0 and v0.13.0 wrote reads back with this release
v0.13.0 Sep 22, 2026
- Zip, Office documents, jars, zlib streams and PNG opened
- JPEG recoded losslessly
v0.12.0 Sep 22, 2026
- Gzip objects compressed further, restored bit for bit
- The store measured at a terabyte
v0.11.2 Sep 21, 2026
- A lost metadata directory rebuilt from the objects
v0.11.1 Sep 21, 2026
- Objects over 64 MB uploaded to S3 in parts
v0.11.0 Sep 21, 2026
- The store speaks S3 itself, with no AWS CLI on the machine
v0.10.2 Sep 21, 2026
- The codec under zstd's licenses: BSD 3-Clause or GPL-2.0
v0.10.1 Sep 21, 2026
- The codec dual-licensed, Apache-2.0 or GPL-2.0
v0.10.0 Sep 21, 2026
- Homebrew, and prebuilt command lines, libraries and Python wheels with every release
- The store a crate of its own, under BUSL 1.1
v0.9.3 Sep 21, 2026
- Python and Go bindings
- The bucket audit and a format specification
- Glyd in lzbench
v0.9.2 Sep 21, 2026
- Faster writes at --max, the corpus suite run again on AWS
v0.9.1 Sep 21, 2026
- The store on S3, and rebase
v0.9.0 Sep 21, 2026
- The store complete: delete, compact, verify, faster puts
v0.8.1 Sep 21, 2026
- The store at scale, with packs for small objects
v0.8.0 Sep 21, 2026
- The store: compression across objects, not only inside them
- Packs for small objects, and dictionaries for record mode
v0.7.0 Sep 20, 2026
- The cold level, for data kept for years
v0.6.0 Sep 20, 2026
- Logs of any shape as columns
- Versions found wherever their content moved
v0.5.0 Sep 20, 2026
- Base mode: a new version stored as its change against the old one
v0.4.0 Sep 20, 2026
- Repeats found far apart in a file at the max and ultra levels
- Record mode: lines, dumps and logs as typed columns
- Dictionaries for small objects
v0.3.0 Sep 19, 2026
- Format v8
v0.2.0 Sep 19, 2026
- The ultra level (-19), for a better ratio
v0.1.0 Sep 19, 2026
- The first public release: the default, fast, turbo and max levels
- Compression and decompression on every core; a C ABI and the command line