werk1112

Getting started

This guide installs the Werk1112 binary, adds a model to the managed store and runs a first local request. Models, accelerator drivers and optional inference runtimes are installed separately.

Install a release binary

Linux and macOS:

sh -c "$(curl -fsSL https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/install.sh)"

The shell installer selects linux-x86_64, linux-x86_64-amd-strix-halo, linux-aarch64-dgx-spark, or macos-aarch64 from the current host. On Linux x86_64 it selects the Strix Halo profile only when specific CPU, DMI, Radeon 8050S/8060S/8040S, or gfx1151 ROCm signals identify the host; other x86_64 systems keep the generic artifact. The ARM64 artifact is accepted only when /proc/device-tree/model or nvidia-smi identifies DGX Spark/GB10; other Linux ARM64 hosts must build from source. The installer verifies the published .sha256 file before extracting any release archive. Models and accelerator runtimes remain separate from the Werk binary. See the Strix Halo and DGX Spark platform guides.

The default destination is $HOME/.local/bin. Select a release or destination with environment variables:

WERK_VERSION=1.10.2 \
WERK_INSTALL_DIR="$HOME/bin" \
sh -c "$(curl -fsSL https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/install.sh)"

Windows PowerShell:

irm https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/install.ps1 | iex

The default Windows destination is %LOCALAPPDATA%\Programs\Werk1112\bin. Select a release, destination or PATH update before invoking the installer:

$env:WERK_VERSION = "1.10.2"
$env:WERK_INSTALL_DIR = "$HOME\bin"
$env:WERK_ADD_TO_PATH = "1"
irm https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/install.ps1 | iex

Verify the installed command:

werk --version
werk --help

Install from source

A source install requires the stable Rust toolchain and the native build dependencies of the selected Cargo features:

cargo +stable install --path . --locked

For a portable CPU-only development build:

cargo +stable build --locked --no-default-features

Target release builds and their platform prerequisites are documented in Building from source.

Add a model

Import copies a local file or directory into Werk’s managed store:

werk import /absolute/path/to/model --name local-model

Pull downloads a Hugging Face repository. Real model repositories require Git and Git LFS:

git lfs install
werk pull organization/repository --name model-name

Repositories containing many GGUF quantizations can be limited to one file:

werk pull organization/gguf-repository \
  --file model.Q4_K_M.gguf \
  --name model-q4

For gated repositories, accept the conditions on the model page and then use one of the supported authentication sources:

werk auth huggingface login
werk auth huggingface status

Werk also recognizes HF_TOKEN, HUGGING_FACE_HUB_TOKEN and the standard Hugging Face CLI token cache. It cannot accept repository conditions for the user.

Inspect what was installed before inference:

werk list
werk inspect model-name
werk doctor --model model-name --debug

An imported model can be catalogued even if no currently installed backend can execute its architecture. doctor and the first cold model load distinguish those cases.

Run inference

Text chat:

werk chat local-model --max-tokens 128

Typed media:

werk image generate IMAGE_MODEL --prompt "A quiet orbital greenhouse"
werk video generate VIDEO_MODEL --prompt "Clouds crossing a mountain ridge"
werk audio generate speech TTS_MODEL \
  --text "Werk elf zwölf ist bereit." \
  --output speech.wav

Use --verbose for measured timings and --debug for request resolution, candidate rejection reasons and the selected runtime:

werk audio generate speech TTS_MODEL \
  --text "Routing diagnostic." \
  --backend auto --verbose --debug

See Media inference for canonical tasks and parameters, and Backends for runtime installation and fallback behavior.

Start the HTTP service

Authentication is enabled by default. Generate a key and start the server:

werk auth api-key generate
export WERK_API_KEY="replace-with-generated-key"
werk serve --model local-model

OpenAI-compatible clients use this base URL:

http://127.0.0.1:11434/v1

For a deliberately unauthenticated loopback development server:

werk serve --model local-model --allow-unauthenticated

Do not expose an unauthenticated server on a public interface. The complete route, authentication and response contract is in the HTTP API.

Uninstall Werk

Linux and macOS:

sh -c "$(curl -fsSL https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/uninstall.sh)"

Windows PowerShell:

irm https://raw.githubusercontent.com/philipbodenbach/werk1112/main/scripts/uninstall.ps1 | iex

The uninstallers remove the binary from the same default or WERK_INSTALL_DIR location. If a model store or API-key file is found, the scripts ask separately before deleting it. The default answer keeps user data.

Managed backend environments have a separate lifecycle. The current cleanup procedure and the absence of a werk backend uninstall command are documented under Backend uninstall and cleanup.

This site is open source. Improve this page.