surya-koritala/glyd v0.32.1 is a release that improves token generation speed for Gemma 4. The GitHub release, published on October 17, 2026, contains speed improvements for the RTX 4080 SUPER plus Gemma 4 12B combination, and a fix for glyd fetch and glyd verify incorrectly passing the model to be downloaded and run. The source of these changes is the v0.32.1 release of surya-koritala/glyd.
Glyd is a project that advertises lossless AI compression. The repository description states that it reduces GPU memory usage by 33% while keeping results bit-identical. The speed comparison uses an RTX 4080 SUPER and Gemma 4 12B without a draft model. Tokens per second changed from 81.2 to 88.9 in chat, from 74.7 to 82.1 after a 4K prompt, and from 59.5 to 64.0 after a 30K prompt. The release explains that output is identical whether or not drafting is used.
The scope demonstrated by these results is limited. Supported environments are Linux with NVIDIA RTX 30 series or newer, and the processed workload is single text requests. So this speed improvement should be interpreted within the GPU and model conditions stated in the release and the supported scope; it is not a result that proves performance in other environments or request types.