Skip to content
KoboldCpp
GitHub

Install on Linux, desktop or headless

KoboldCpp for Linux is a single file. There is no installer and no package to add.

  1. Pick your file. See Download KoboldCpp if you are unsure.

    Your PCFile
    NVIDIA graphics cardkoboldcpp-linux-x64
    AMD, Intel or no graphics cardkoboldcpp-linux-x64-nocuda
    Older processor or older NVIDIA cardkoboldcpp-linux-x64-oldpc
  2. Download it with your browser. It usually lands in ~/Downloads.

  3. In a terminal, go to the folder with the file, make it executable and run it. This example uses the NVIDIA file; replace the name with yours:

    Terminal
    cd ~/Downloads
    chmod +x koboldcpp-linux-x64
    ./koboldcpp-linux-x64

    The launcher opens.

Continue with Choose a model.

To download from the terminal instead, for example on a server:

Terminal
curl -fLO https://github.com/LostRuins/koboldcpp/releases/latest/download/koboldcpp-linux-x64

"Permission denied" means the file is not executable yet; run chmod +x on it. If the main build fails on an older system, try koboldcpp-linux-x64-oldpc, the build with the greatest compatibility.

ProgramNeeded for
yad or zenityA better file picker. yad is used if both are installed; without either, the launcher uses its built-in picker.
vulkan-toolsDetecting your graphics card and its memory with vulkaninfo. NVIDIA cards are also detected with nvidia-smi, and AMD cards with rocminfo if ROCm is installed. If none of these works, you see "Unable to detect VRAM."
xdg-openOpening the browser after the model has loaded.
  • The file picker does nothing (often on Wayland). Install yad, or turn on Use Classic FilePicker on the launcher's Extra tab. The console may show Zenity/YAD sanity check failed.
  • "Warning, GUI failed to start". The launcher could not open a window, for example because there is no display. KoboldCpp then tells you to use the command line. See Headless use.
  • You want to see the console output. Spawn Terminal on the Extra tab opens an xterm window that mirrors it.

More: OS-specific problems.

Start with the Vulkan backend in koboldcpp-linux-x64-nocuda. Install vulkan-tools first, so the launcher can detect your card and select Use Vulkan by itself. Without it, KoboldCpp does not find the card and runs on the CPU.

If generation is slow with Vulkan, try the experimental official ROCm build. It does not need ROCm installed. See ROCm build for AMD on Linux and GPU backends: ROCm.

With a model on the command line, KoboldCpp skips the launcher and needs no display:

Terminal
./koboldcpp-linux-x64 --model mymodel.gguf
  • ./koboldcpp-linux-x64 --help lists all flags.
  • --config mysettings.kcpps starts with a saved configuration. A config that contains a model also skips the launcher.
  • --skiplauncher overrides --showgui, including a showgui saved in a config. It does not pick a model: without a model from the command line or a config, KoboldCpp exits with an error.
  • By default KoboldCpp listens on all network interfaces, so other devices on your network can reach it. Add --host 127.0.0.1 to allow connections from this machine only.

More: Command line, Remote access, Passwords and security.

There is no ready-made file for ARM Linux. Build from source; koboldcpp.sh supports aarch64.

The x64 files need glibc 2.29 or newer, for example Ubuntu 20.04 or newer. ldd --version shows your glibc version.

The community maintains packages that are unofficial and get no official support: koboldcpp-cuda and koboldcpp-hipblas in the AUR, and koboldcpp in Nixpkgs.