The user path

For someone who wants ignis running and answering, on a card it supports, with its defaults. Nothing on this path asks you to choose a setting: the defaults are the ones measured on the owner’s own cards.

1. What ignis is — the models it serves, the cards and systems it runs on, and what it deliberately does not do; then Why ignis, what sets it apart, with its numbers.
2. Install — the container image or a release archive, and what each needs from the host: the NVIDIA driver, and for the archives the CUDA runtime. Or let the coding agent you already use do it: Install with a coding agent.
3. Quick start — one command to a running server, the Playground in the browser, one request with curl.
4. Get a model — the built-in catalog, ignis-server model download, and pointing at an artifact you already have.
5. The Playground — every view, with screenshots: the chat and its settings, subagents and tools, Decide, the image studio, Transcribe, the Monitor and the Server view.
6. Connect a client — the base URL and the key, the OpenAI SDKs, coding agents and chat front ends.
7. Troubleshooting — what a refusal or a WARN line means, and what to do about it.
8. Measure a 16 GB card — on an RTX 5060 Ti 16 GB, 5070 Ti or 5080, one command writes a record of what the server does there, to paste into a GitHub issue.

Tip

The server holds one model at a time and gives it the whole card. A client that names two different models in turn makes the server reload back and forth; point every request at the same id (Connect a client).

Tip

The first start after a download, or the first request after the server unloaded an idle model, waits for the load. Later requests do not (Idle unload).

When the defaults stop being enough — a longer context, a smaller card, a machine without internet — continue on the advanced path.