Note
For easier repo installation, you can now use extrepo.
See the documentation for details.
Note
The repo URL has moved from ollama-repo.jason-9eb.workers.dev to packages.lingfish.net.
The old URL will continue to work, but new installations should use the new URL.
See the documentation for details.
Note
Whilst things were broken with the repo, I took the time to change some things:
- You will need to update your
sources.listto change frommaintonon-free. rcbuilds are now in their own suite too, so there's no need anymore for previousaptpinning instructions (feel free, and I recommend, removing the pin file if you previously had one).- I will only be hosting the latest versions of released and
rcpackages. - As of 20 April, 2026, builds (and hence packages) were broken. This is due to the ollama team now pushing builds (specifically nvidia libraries) that are greater than 2GB in size, which breaks GitHub release upload limits. Therefore, packages cannot be uploaded to GitHub releases anymore, due to their size.
Please see the documentation for updated info.
This is a GitHub workflow to build Debian packages of ollama based on tags.
It installs the same files as the binary tarballs that you can download from their releases page, plus more.
It will also install, enable and start a service (systemd on systemd systems, init.d on sysvinit/OpenRC), as well as add a system user to run ollama. It also adds the
source documentation.
amd64 and arm64 architectures are packaged.
A CHANGELOG is maintained tracking packaging changes.
Click here to get instructions on repo usage.
Note
The sources.list codename has changed from Debian distro names (like bookworm) to stable. Ollama themselves don't
specify what releases/distros they support, so I've changed to using an agnostic one too. bookworm will remain for a
while, and then be removed.
The shared libraries that the ollama team put into their binary tarballs are now split into three packages:
libollama-common(CPU dispatch + Vulkan) — required byollamalibollama-nvidia(CUDA backend) — optionallibollama-amd(ROCm backend) — optional
CPU inference works out of the box. GPU packages are additive.
For Vulkan GPU acceleration on Linux, a Vulkan ICD driver must be installed separately. See Vulkan Support in the documentation.
