Glyd v0.29.0: Qwen3.5, and vLLM compiled by default

Glyd v0.29.0 is out. It is the first release measured on Qwen3.5, it starts vLLM compiled, it ships Glyd’s own engine as an opt-in, and it puts all of Glyd under one license. Install or update with the same line:

curl -LsSf https://getglyd.com/install.sh | sh
glyd run Qwen/Qwen3.5-9B

Qwen3.5, measured

Qwen3.5-9B under vLLM 0.30.0, the weights as vLLM’s log gives them, against vLLM’s own bf16 on the same GPU (the run’s record):

On the 24 GB RTX 4090, bf16 did not start with vLLM’s default settings: it ran out of memory at a 4,096-token context. glyd run runs the model there with a 189,440-token context.

Qwen3.5 is the model glyd run and glyd doctor now name and suggest, and vLLM gets Qwen3.5’s own tool-call and thinking parsers: the thinking comes apart from the answer, and Open WebUI’s chats with tools are answered.

vLLM compiled by default

glyd run and glyd serve start vLLM compiled, vLLM’s own default, wherever the chat keeps 8,192 tokens of context, and eager only where that gives a longer chat; v0.28 always started it eager.

A first start compiles, and takes longer: Qwen3.5-9B was up in 135 s on the RTX 5090, against 72 s eager.

Glyd’s own engine, opt-in

The glyd-gpu wheels now carry Glyd’s own engine. It is off by default: every model runs on vLLM, as in v0.28, and the first line a command prints says which runs the model. GLYD_ENGINE=1 asks for the engine; it does not run 2026 models yet, so they run on vLLM. Serving with vLLM has the details.

One license

From v0.29.0 all of Glyd is under the Business Source License 1.1: free forever for personal and non-commercial use on computers you own or rent for yourself, and for anyone to try, test and develop with; commercial use needs a license. Releases up to v0.28 keep the licenses they shipped with. The license.

Model files

glyd --max makes model files smaller: safetensors and GGUF. The 21 files cut from 2026 models in the record, Qwen3.8, Gemma 4, GLM-5.3-Flash, MiniMax-M3 and Mistral Small 4 (3.72 GB in all), are 32.5% smaller as a set, against 24.7% with v0.28.0’s --max, 25.4% with zstd -19 and 27.3% with xz -9, every file read back bit for bit; the set’s numbers are computed from the record’s table, each file weighted by its size (table and logs). Files and streams of every earlier release read as before.

Not there yet

Qwen3.5-9B’s weights are 28.2% smaller on the RTX 4090 and 21.3% on the RTX 5090: short of 33%. Parts of the newest models are not compressed yet (the vision tower, the multi-token-prediction head and the linear-attention layers), and the RTX 5090 takes a larger layout by default because it is faster there. 33% on every 2026 bf16 model is next on the roadmap. Every gap, with its logs: the known gaps.

Every change, with its numbers: the changelog and the release notes.