# Ollama with a preloaded model

An [`ollama/ollama:latest`](https://hub.docker.com/r/ollama/ollama) image with the following model downloaded during the build:

```text
orcarouter/Qwen3.8-27B-Uncensored:q3_K_L
```

The image is published to the public Gitea container registry as:

```text
gt.uo1.net/1movmedia/ollama-with-model:latest
```

## Run

```sh
docker pull gt.uo1.net/1movmedia/ollama-with-model:latest
docker run --rm --gpus all -p 11434:11434 \
  gt.uo1.net/1movmedia/ollama-with-model:latest
```

The Ollama API is then available at `http://localhost:11434`. Start a request with the preloaded model:

```sh
curl http://localhost:11434/api/generate \
  -d '{"model":"orcarouter/Qwen3.8-27B-Uncensored:q3_K_L","prompt":"Hello"}'
```

Remove `--gpus all` to run without an NVIDIA GPU.

## Build locally

The base image and model have defaults and can be overridden with Docker build arguments:

```sh
docker build -t ollama-with-model:latest .

docker build -t ollama-with-model:custom \
  --build-arg BASE_IMAGE=ollama/ollama:latest \
  --build-arg MODEL=orcarouter/Qwen3.8-27B-Uncensored:q3_K_L \
  .
```

Building downloads the complete model, so the runner needs enough disk space for the Ollama base image, model data, and build layers.

## Publish from Gitea Actions

The workflow in `.gitea/workflows/publish.yml` builds the image by digest, then publishes the `latest` tag in a separate job. After a successful publication, it removes every other version of the image from the Gitea registry. It runs on pushes to `main` and can also be started manually with `workflow_dispatch`.

Configure the following repository Actions variable and secret before running it:

- Variable `REGISTRY_USERNAME`: a Gitea user with write access to the `1movmedia` package registry.
- Secret `REGISTRY_TOKEN`: a Gitea personal access token with package read/write permission.

The `1movmedia` organization must be public for unauthenticated image pulls.