Release · Oct 9, 2026 · Glyd
Glyd v0.29.4: levels, and a new one, swift
Glyd v0.29.4: glyd run MODEL:LEVEL picks the bits the weights keep. New :swift, about 5.5 bits per weight: Qwen3.8-27B in 17.4 GiB at 47.5 tokens a second.
Glyd v0.29.4 is out. glyd run MODEL:LEVEL now sets how many bits the weights keep on the GPU, and there is a new level.
:penguinis lossless, the default.:kestrelis about 6.5 bits per weight. It is v0.29.3’s--lossy, which still works.:swiftis new, about 5.5 bits per weight.
curl -LsSf https://getglyd.com/install.sh | sh
glyd run Qwen/Qwen3.8-27B:swiftKestrel and swift are not lossless. They run on Glyd’s own engine only (Qwen3.5, Qwen3.6 and Qwen3.8 dense models), and a level that is not built yet is refused with the list of those that are. glyd --help and the READMEs show the levels.
Qwen/Qwen3.8-27B, a prompt of 1,024 tokens and 256 more, one user, one A100 80GB:
| Level | Weights on the GPU | Tokens a second | GPU memory in use |
|---|---|---|---|
:swift |
17.4 GiB | 47.5 | 19.5 GiB |
:kestrel |
20.5 GiB | 46.3 | 22.6 GiB |
:kestrel (and --lossy) is closer to bf16 than before. With Qwen/Qwen3.8-27B the next token’s distribution differs from bf16’s by a KL divergence of 0.00477 on 253,952 positions of WikiText-2, against 0.00513 for v0.29.3 and 0.00465 for Qwen/Qwen3.8-27B-FP8. :swift is 0.01393. Lower is closer.
The trade-off is the start. A lossy level fits the weights more closely while they load, so starting takes longer: on the A100, 127 s for kestrel and 157 s for swift (v0.29.3: about 50 s).
Full details are in the release notes and the changelog; what is not done yet is on the known gaps.