Docs · Use Nodeau

The Playground

Start a model, open the Playground, and use it. Everything happens on your own machine, and every answer shows which machine and GPU produced it and how fast.

Open it#

On the machine where you run nodeau, which is the control-plane machine of a fleet, or your Mac:

nodeau playground                 # every model running here
nodeau playground qwen-local      # with that model already chosen

It prints an address that signs your browser in. Open it in your browser on the same machine; the sign-in lasts until the Playground stops. The Playground is served by the same local server as nodeau dashboard, on loopback only, so if the dashboard is already running, nodeau playground prints the Playground's address on it and you're done. Press Ctrl+C to stop it.

The list on the left shows the models that are running here, never the whole catalog, because a model has to be running to answer. Each one says what it does (chat, images, embeddings or reranking), which machine it's on and which GPU it holds. A model that can't take a request right now is listed with the reason and what to do about it.

Nothing running yet? Start something first:

nodeau run qwen3.5-4b-q4km
nodeau playground

Chat#

Type a message and press Send, or Ctrl+Enter (Cmd+Enter on a Mac). The answer streams in as the model writes it, and Markdown and code are rendered for reading, with a copy button on every code block. Stop ends the answer there and frees the model straight away. Retry asks the last question again and replaces the answer it gave. A reasoning model's thinking is kept apart from its answer, folded away until you open it.

Settings holds only what a request can change: temperature, max tokens, top-p, a seed, and whether to stream. Leave a field empty for the runtime's own default. Context size and parallel requests belong to the running model and are set when it starts, so they aren't offered here.

If you set Max tokens and a reasoning model spends them all on thinking, the Playground tells you, so you can give it more room.

Images#

With a model that reads images, such as gemma-4-e4b-qat-q4-0, the chat has a place to add an image: choose a PNG, JPEG or WebP up to 10 MB, or drop one on it. A WebP is redrawn as a PNG on the page, because that's what the model's runtime reads. The image goes from your browser to Nodeau on your machine and to the model, and is held only for that request. Models that don't read images don't offer it.

Embeddings#

Add one or more texts and press Create embeddings. For each text you see how many dimensions the vector has, its first few values and its length. Copy the full JSON copies every value exactly as the model returned it, ready to paste into your own code.

Reranking#

Write a query, add the passages you want ranked, and press Rank passages. They come back in the order of the model's own scores, highest first, with each score shown beside its passage.

Batch jobs run from the command line; see batch inference.

Where the answer came from#

Beside every model the Playground shows where it runs: the machine, the GPU, and on Linux the memory Nodeau predicted for it when it admitted it. Why this GPU? shows the placement Nodeau recorded: the machine and card it chose, the rule that decided, the scheduling mode, and the speed it predicted.

Under every answer, each figure says who measured it:

A figure nobody reported is shown as missing, never as zero.

Inspect and copy#

Inspect the request and the response shows the exact request the page sent and the response it got back, the model's own endpoint and the HTTP status. A streamed answer is shown assembled into one response.

Copy this call as code gives you the same request as cURL or Python, aimed at the model's own endpoint, ready to run on this machine. On Linux it fills the key the endpoint checks with:

export NODEAU_API_KEY="$(nodeau auth token qwen-local)"

On a Mac the endpoint answers only programs on the Mac itself and checks no key, so the snippet has none. Python uses the official OpenAI client for chat and embeddings; reranking uses a plain HTTP request.

Compare#

Compare sends one request to two or more running models at once and shows the answers side by side, each with its own machine, GPU and figures. Choose models that do the same kind of work: chat models together, or embedding models together. An image is offered only when every chosen model reads images. Each column has its own Stop, and stopping one leaves the others to finish. The Playground shows what each model did and leaves the judging to you.

Every model you compare has to be running at the same time, so how many you can put side by side depends on what your machines can hold at once. On a Mac, that is the Mac's unified memory, shared with everything else it's doing.

A preset saves the chosen models, the system prompt and the settings in this browser, so a comparison you run often is one click away. A preset never keeps a prompt or an answer.

Your prompts stay on your machine#

• The page is served by Nodeau on your machine, and your browser only ever talks to it. The page's content security policy holds it there.
• Nodeau sends each request to the chosen model's own endpoint on the same machine, and before it does, it checks that the endpoint really is that model's.
• Nodeau attaches your API key itself. Your browser never holds it, and it never appears in a copied snippet.
• No conversation is kept: closing the tab ends it.
• Nodeau Cloud never sees a prompt, an answer or an image.

From your account#

In your account at app.nodeau.ai, a running model on a connected fleet has Open in Playground, which tells you the command to run and on which machine. The Playground itself always runs on that machine, so your prompts stay there.

If something goes wrong#

These docs describe the current release, the beta channel. Spotted something wrong or missing? Tell us. A report from a machine we have never seen is the most useful thing we get.