Prebuilt command-line tools let a local AI setup run without a compiler or build toolchain. The official llama.cpp release archives contain ready-built executables for installation on Windows 11 and Ubuntu Linux by extracting them into a user-owned folder.

The archive must match the operating system, processor architecture, and backend. A CPU build is suitable when GPU acceleration is unnecessary; a CUDA, Vulkan, or other accelerator build also needs its matching hardware and runtime libraries. Windows CPU builds require the matching Microsoft Visual C++ v14 Redistributable, which may need administrator approval to install.

An extracted release stays in its own folder and does not add itself to PATH or update automatically. The executables and their bundled libraries belong together. Model weights are a separate download, so checking the server's version and device list needs no GGUF file and does not start an HTTP listener.

Steps to install llama.cpp from prebuilt release binaries:

Follow the section for your operating system. Both use the official CPU archives from release b12328; a different release needs its own tag and matching asset name from the release list. The Linux package commands apply to Ubuntu 26.04.

Windows 11

  1. Check the Windows processor architecture in a normal PowerShell session.
    $ [System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture
    Arm64

    Arm64 uses win-cpu-arm64.zip; X64 uses win-cpu-x64.zip. These are different native builds.

  2. Set the release tag and CPU archive name for your architecture.
    $ $release = "b12328"
    $ $asset = "llama-$release-bin-win-cpu-arm64.zip"

    For X64, replace arm64 with x64 in the asset name. A missing VCRUNTIME140.dll or MSVCP140.dll is a runtime dependency error, not a reason to download individual DLLs from other sites.

  3. Have an administrator install the architecture-matched Microsoft Visual C++ v14 Redistributable if it is missing.

    The runtime installer changes system libraries and needs administrator approval. The official ARM64 installer matches the ARM64 archive; an X64 archive needs the X64 runtime. Its publisher must be Microsoft Corporation. The release download and extraction use your normal account.

  4. Create a new folder for this release under your local application data.
    $ New-Item -ItemType Directory -Path "$env:LOCALAPPDATA\llama.cpp\$release"

    A separate release folder keeps older binaries and libraries from being mixed into this installation.

  5. Change to the release folder.
    $ Set-Location "$env:LOCALAPPDATA\llama.cpp\$release"
  6. Download the selected archive from the official repository.
    $ Invoke-WebRequest -UseBasicParsing "https://github.com/ggml-org/llama.cpp/releases/download/$release/$asset" -OutFile $asset
  7. Extract the archive into the release folder.
    $ Expand-Archive -Path $asset -DestinationPath .

    The Windows archive places llama-server.exe beside its DLLs. Moving only the executable elsewhere can prevent it from starting.

  8. Confirm that the extracted server reports build 12328. Example output:
    $ .\llama-server.exe --version
    version: 0.9.0-dev (build 12328, commit 3bce8bd50)
    built with Clang 20.1.8 for Windows arm64

    The build number must match the selected b12328 archive. An incompatible-binary or missing-DLL error means the installation is not ready.

  9. Check that the server can enumerate devices without an error. Example output:
    $ .\llama-server.exe --list-devices
    Available devices:
      (none)

    A CPU archive can report Available devices: followed by (none). That means no accelerator is listed; it does not mean the CPU build is unusable. A GPU archive needs its intended accelerator to appear before GPU use can be assumed.

Ubuntu Linux

  1. Identify the Linux processor architecture.
    $ uname -m
    aarch64

    aarch64 uses ubuntu-arm64.tar.gz; x86_64 uses ubuntu-x64.tar.gz. The ubuntu CPU archives differ from the CUDA, Vulkan, and Snapdragon-specific assets.

  2. Refresh the Ubuntu package index.
    $ sudo apt-get update
  3. Install the download tool and runtime libraries.
    $ sudo apt-get install curl ca-certificates libgomp1 libstdc++6 libssl3t64

    libgomp1 supplies the OpenMP runtime, libstdc++6 supplies the C++ runtime, and libssl3t64 supplies OpenSSL 3. The archive supplies the llama.cpp libraries.
    Related: Install cURL on Ubuntu

  4. Set the release tag and CPU archive name for your architecture.
    $ release=b12328
    $ asset="llama-$release-bin-ubuntu-arm64.tar.gz"

    For x86_64, replace arm64 with x64 in the asset name.

  5. Create a user-owned installation directory.
    $ mkdir -p "$HOME/llama.cpp"
  6. Change to the installation directory.
    $ cd "$HOME/llama.cpp"
  7. Download the selected archive from the official repository.
    $ curl --fail --location --remote-name "https://github.com/ggml-org/llama.cpp/releases/download/$release/$asset"

    --location follows GitHub's asset redirect, and --fail reports an HTTP download failure.

  8. Extract the release and its bundled libraries.
    $ tar -xzf "$asset"

    The archive creates llama-b12328. Its shared libraries must remain with the executables; a compatible system runtime is still required.

  9. Change to the extracted release directory.
    $ cd "llama-$release"
  10. Confirm that the extracted server reports build 12328. Example output:
    $ ./llama-server --version
    version: 0.9.0-dev (build 12328, commit 3bce8bd50)
    built with GNU 14.2.0 for Linux aarch64

    The architecture and compiler line varies by archive. A missing-library or GLIBC version error means the system libraries are incompatible with that build.

  11. Check that the server can enumerate devices without an error. Example output:
    $ ./llama-server --list-devices
    Available devices:
      (none)

    (none) is normal for this CPU archive. The version and device commands both exit successfully without loading a model. Future sessions can run the tools by their full path or from this extracted directory.

Share this guide