Skip to main content

Release Notes

On this page

This section records what each GPUWorks release adds, changes, fixes, and secures, newest first. For upgrade steps, see Upgrades and Version Compatibility.

Versions

VersionRelease dateTypeHighlights
v26.11.32026-11-12Patch releaseHygon cards reporting only BW match built-in templates, Hygon cards other than card 0 work under Docker/nerdctl, deployments no longer refused when the selected model and weight source disagree
v26.11.22026-11-08Patch releasePrivate-network sources work by default, storage paths no longer limited to default directories, local models deploy with inference templates, benchmark metrics corrected for reasoning models, Go 1.26.10
v26.11.12026-11-04Patch releaseDefault vLLM and vLLM-Omni runners move to r4, Qwen3-0.6B deploys on V100, about 20 templates fixed from A100 runs, archives ship third-party license texts
v26.11.02026-11-03First releaseFirst version of the 26.11 release line

Version Numbers

Releases use calendar versions vYY.M.PATCH: YY.M is the year and month the release line was cut from master, and PATCH starts at 0. Patch releases on a line are mostly bug fixes; a fix is merged into master first and then backported to the release branch. The "Upgrade notes" section of each version page lists behavior changes to check. For the full rules on version numbers and publishing, see Release and Embedded Assets.

Downloading and Verifying

The one-line installer get.sh picks the fastest of dl.gpuworks.ai, Alibaba Cloud OSS, and Huawei Cloud OBS and checks the archive against SHA256SUMS before installing; upgrades from the console read the ed25519-signed release index. For a manual download, archives live under https://dl.gpuworks.ai/releases/<version>/, and SHA256SUMS in the same directory lists the checksum of each platform's archive. The download sources keep every release's archives permanently and the most recent 30 nightly builds.

Was this page helpful?