Private chat on your own hardware: Marigold as the Open WebUI backend, serving a self-hosted instruct model. This assumes the setup guide is done.
Run
marigold package create marigold-examples/chat -o /tmp
marigold package install /tmp/chat-<version>.tar.gz
marigold cache populate chat
marigold application start chat
Open http://localhost:3000 and chat with the model using
Open WebUI.
The package lists webui in its compose_files, so starting it adds
Open WebUI to the platform, as a client of the API. The package’s
application checks that its model is in the catalogue and exits;
marigold application status chat shows exit code 0 when it is.
Open WebUI lists every model in the catalogue, including models cached for other packages.
Models
qwen/qwen3-8b– instruct
This package runs Qwen locally out of the box. To change the model, see adding a new model.
From your own code
Open WebUI reaches the model through Marigold’s OpenAI-compatible
endpoints. Any OpenAI SDK client can do the same at
http://localhost:8000/v1; the
setup guide has a
Python example, and
Cursor and VS Code covers editors.
Stop
marigold application stop chat # the application container
marigold platform stop --applications # everything, Open WebUI included
Open WebUI runs with the platform, so stopping the application leaves it running.