Release · Oct 8, 2026 · Glyd
Glyd v0.29.2: Qwen3.8-27B in 21 GiB with glyd run --lossy
Glyd v0.29.2 adds glyd run --lossy: Qwen3.5, 3.6 and 3.8 at about 6.8 bits per weight, 24% less GPU memory for weights than FP8 at its speed.
Glyd v0.29.2 is out. It adds glyd run MODEL --lossy, which runs Qwen3.5, Qwen3.6 and Qwen3.8 dense models on Glyd’s own engine at about 6.8 bits per weight. It is not lossless: every weight is rounded.
curl -LsSf https://getglyd.com/install.sh | sh
glyd run Qwen/Qwen3.8-27B --lossyMeasured on one L40 with Qwen3.8-27B, a prompt of 1,024 tokens and 256 more, one user:
| Weights on the GPU | Tokens a second | First token | |
|---|---|---|---|
Glyd, --lossy |
21.2 GiB | 25.8 | 526 ms |
| vLLM 0.30.0, Qwen’s FP8 checkpoint | 28.0 GiB | 25.4 | 504 ms |
On 13,243 questions of six suites (ARC-Challenge, HellaSwag, WinoGrande, PIQA, BoolQ, MMLU), the same values scored a 78.68% mean, against 78.54% for bf16 and 78.55% for the FP8 checkpoint: a statistically inconclusive difference.
What it is for now: one user at a time, in the terminal (one answer with --prompt, or a chat), on a GPU with 32 GB or more for Qwen3.8-27B. It downloads the model’s bf16 checkpoint and packs it at each start, about 70 s on the L40 once the checkpoint is on disk. Serving several users and an API are not in this release. Without --lossy, glyd run is as before: every weight bit for bit.
Full details are in the release notes and the changelog.