Quick start

You need a machine with a supported NVIDIA card and driver (What ignis is) and a container runtime that can give the container the GPU. The commands use podman; with docker, write --gpus all in place of --device nvidia.com/gpu=all.

1. Run the container

podman run --rm --device nvidia.com/gpu=all -p 8000:8000 \
  -v /path/to/models:/models \
  ghcr.io/gpillon/ignis:latest --download-path /models

/path/to/models is a directory you can keep. The server runs in the image as ignis (uid 10001), so that directory must be writable by uid 10001 for the download to land in it (the image’s user).

With no model in /models, the server fetches the default one, qwen3.8-27b: 18.1 GiB from Hugging Face (on a 16 GB card, the id’s 16 GB entry, qwen3.8-27b-16g, 12.5 GiB), every file checked against its pinned size and SHA-256 before it is used. On a terminal the server first asks ([y/N], on stderr), naming the size, the source and the destination. In a container there is no terminal to answer, so it downloads without asking (Letting the server fetch it). An interrupted download resumes from its .part file.

Once the files are in, the model loads onto the card (a cold load took 13-33 s on a 5090, Idle unload). Until it serves, /health answers 200, and /ready and the /v1 routes answer 503 server_not_ready with Retry-After: 5 (Deadlines).

Tip

To download ahead of time, on a machine with no GPU and no driver: ignis-server model download qwen3.8-27b, with --out <dir> to choose the directory (Fetching one).

You can run a release archive instead of the container: unpack the Windows .zip or the Linux .tar.gz (Install) and run ignis-server with no arguments. It listens on 127.0.0.1:8000 and fetches the model into ./models.

2. Open the Playground

Go to http://127.0.0.1:8000/ui/. The header shows the loaded model; type a message in the chat (The Playground).

3. Send a request

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen3.8-27b","messages":[{"role":"user","content":"Hello"}],"max_tokens":256}'

The answer is a standard chat completion: choices[0].message.content holds the text and usage the token counts. GET /v1/models lists the loaded model and what the server can switch to:

curl http://127.0.0.1:8000/v1/models

If a request returns 503 server_not_ready, the model is still loading: retry after the Retry-After seconds. Other messages are in Troubleshooting.

What the defaults are

• The API and the Playground are on port 8000. There is no API key until you set one (Connect a client).
• The metrics listener (127.0.0.1:9464) stays inside the container; the Playground’s Monitor reads the same data at /ui/metrics on port 8000 (Install shows how to scrape it from outside).
• The default profile is auto: the start picks rtx50-16gb, rtx5090 or rtxpro6000 by the card’s memory (Profiles and defaults).
• The model unloads after 300 s without a request and loads again on the next one (Idle unload).

Next

• Connect a client to use it from an SDK or an agent.
• Get a model for Flash-Next, the image model and your own artifacts.
• Install for a cache directory that keeps the n-gram cache and KV-disk, and for how the container stops.