Roadmap

Now: v0.29.1

Released on 6 October 2026: newer-model startup fixes, more loaded weights compressed, and model correctness fixes. The release post.

v0.29.0

The preceding release and its measured results:

• Qwen3.5, measured. Qwen3.5-9B's weights are 28.2% smaller than bf16's on an RTX 4090 and 21.3% on an RTX 5090; on the 24 GB RTX 4090, where bf16 did not start with vLLM's default settings, glyd run runs it with a 189,440-token context.
• vLLM compiled by default. glyd run and glyd serve start vLLM compiled wherever the chat keeps 8,192 tokens of context.
• Glyd's own engine ships, off unless you ask for it; every model runs on vLLM by default.
• All of Glyd under the Business Source License 1.1, and model files (safetensors and GGUF) smaller with glyd --max.

What changed, with the measured numbers: the v0.29.0 post and the changelog.

Next: quantized checkpoints, then one-third savings across GPUs

• Quantized checkpoints. Lossless compression for FP8, FP4 and integer-quantized models, with original values preserved.
• 33% on every 2026 bf16 model, every maker. The parts not compressed yet: vision towers, multi-token-prediction heads and the linear-attention layers.
• One-third savings by default on more GPUs. RTX 5090 first, then A100 and H100, with memory and performance measured before defaults change.
• Lossless FP8 checkpoints, such as GLM-5.3, DeepSeek-V4, Mistral Small 4 and Mistral Medium 3.5, smaller in GPU memory with every value exact.

Later: Glyd's own engine

• Glyd's own engine runs Qwen3.5 and Gemma 4, then the mixtures of experts, and takes over from vLLM wherever it is faster at every load, from one user to 256, with less memory. vLLM stays the fallback for the rest.
• The lossless KV cache for the newest attention: linear attention and a head size of 256, as in Qwen3.5 and Qwen3.8. Today these keep vLLM's own KV cache.
• Apple Silicon Macs.

Every item ships on a measured result, taken against bf16 on the same machine. The numbers come with their logs on the benchmarks page.

How releases are supported

• Order, not dates. Releases ship in the order above when their checks pass. Each one is tested before it ships by installing it the way users do, on real GPUs.
• Fixes go into the latest release. Upgrading is the fix: run the install line again, or pip install -U "glyd[vllm]". Earlier releases stay installable from PyPI and from their tags on GitHub.
• Patch releases, such as 0.29.1, carry fixes between the releases above.
• Compressed files stay readable. Files compressed by the glyd program with any earlier release keep decoding with every later one.
• vLLM versions. Each release supports the vLLM version it was tested with; v0.29.1 supports vLLM 0.30, and glyd[vllm] installs it. A newer vLLM is supported in a later Glyd release once it passes the same checks. With another version installed, the plugin loads nothing and says why.
• Before 1.0, options and defaults can change between releases. Every change is listed in the changelog.
• Security reports go privately through GitHub's Security tab. How, and what counts: SECURITY.md.

What Glyd runs on

The codec's own research plan, item by item: ROADMAP.md.