Release · Oct 9, 2026 · Glyd
Glyd v0.29.3: glyd run --lossy, faster and closer to bf16
Glyd v0.29.3: glyd run --lossy at about 6.5 bits per weight, closer to bf16, and Qwen3.8-27B at 31.3 tokens a second on one L40 against FP8's 25.4.
Glyd v0.29.3 is out. glyd run MODEL --lossy now runs Qwen3.5, Qwen3.6 and Qwen3.8 dense models at about 6.5 bits per weight (6.8 in v0.29.2), closer to bf16 and faster. It is still not lossless: every weight is rounded.
curl -LsSf https://getglyd.com/install.sh | sh
glyd run Qwen/Qwen3.8-27B --lossyQwen3.8-27B, a prompt of 1,024 tokens and 256 more, one user, in tokens a second:
| Weights on the GPU | One L40 | One A100 80GB | |
|---|---|---|---|
Glyd v0.29.3, --lossy |
20.5 GiB | 31.3 | 45.0 |
Glyd v0.29.2, --lossy |
21.2 GiB | 25.8 | 18.5 |
| vLLM 0.30.0, Qwen’s FP8 checkpoint | 28.0 GiB | 25.4 | 47.5 |
vLLM’s L40 figure was measured on another L40; on the A100 both ran on the same GPU.
How close it stays to bf16: over 253,952 positions of WikiText-2, the next token’s distribution differs from bf16’s by a KL divergence of 0.00513, against 0.00831 for v0.29.2 and 0.00465 for the FP8 checkpoint. Lower is closer.
It starts faster as well: packing the weights at each start takes about 52 s on the L40, against 78 s for v0.29.2. The rest is as in v0.29.2: one user at a time, in the terminal (one answer with --prompt, or a chat), and a GPU with 32 GB or more for Qwen3.8-27B. Without --lossy, glyd run is as before: every weight bit for bit.
Full details are in the release notes and the changelog.