# Playground (Preview)

> Try models and configure compute jobs, agents and sandboxes in one workspace.

Source: https://www.nodus-compute.ai/docs/guides/inference/playground/
Build revision: 718a07fed53e8ac579fe009e9513f452d8de8e98

**Alpha:** this feature may change.

Sign in to the console and open **Inference → Playground**. Models, Playground and Endpoints are views within Inference; older Playground links redirect there. Choose your organization and project before starting. Switching a mode does not launch work or spend credits.

The **Models** view offers a compact list and a card grid, with model logos and customer prices. Click anywhere on a model tile to open its Playground. The sidebar has collapsible input and output modality, context length, capability, input-price and lab filters. Labs show author logos and model counts, largest first; **Show all** expands the full list. Input prices are per million tokens and exclude models priced only by audio duration or characters. **Reset refinements** clears the sidebar selections.

The catalog includes only models with reviewed open-weight evidence and a supported, priced serving route. Choose **All models**, **Chat**, **Decisions**, **Audio** or **Embeddings**. For All models and Chat, the **Capabilities** section requires every selected capability: Vision, Tools and Reasoning. Changing model type clears these capability filters. Sort by name, context or lowest input-token price.

Search the catalog, then choose **Try** or **Open in Playground** on a model. Switch between **Models** and **Playground** at the top without losing your search, model type, selected model or draft. **Clear filters** returns to the full catalog. Older links with multiple model types open All models.

|Mode|What you can do|
|-|-|
|Chat|Have a conversation, stream replies, adjust settings and inspect usage details.|
|Decisions|Build typed questions and compare up to three available models on identical input.|
|Audio|Generate speech or upload an audio file for transcription, depending on the selected model.|
|Embeddings|Turn text into a vector and inspect the returned JSON.|
|Code|Configure a container command as a Job, review resources and cost, then launch. CPU is selected initially.|
|Agents|Choose a saved agent, describe its task, review its model spending limit and start a run.|
|Sandboxes|Configure an isolated environment, review its resources and create it.|

Available model choices come from your organization’s catalog. If a model mode has no available models, the Playground says so.

**Chat** keeps earlier user messages and successful replies in the next request. Enter sends; Shift+Enter adds a line. **New chat** clears the conversation, and choosing a different model starts fresh. Conversation history stays in memory for this view; it is not saved across page reloads. **Settings** holds the system prompt, temperature and output cap. Open **Usage & details** below a reply for its receipt.

A local console in **Preview mode** explicitly labels simulated responses. It calls no model and charges no credits; its sample output and catalog are not live serving evidence.

Model requests use your console session and project. Review the displayed catalog rate before sending. Receipts may take time to settle; **Cost pending** does not mean free. Stopping a model request locally does not guarantee that an already-dispatched request will not finish or be billed.

Audio uploads must be no larger than 25 MiB. Transcription is billed for at least ten seconds per request. Speech uses the model’s default voice unless you supply a supported voice name. Generated audio can be played and downloaded. Embedding previews show at most 32,000 characters.

Code, Agent and Sandbox modes reuse the console’s existing configuration and review flows. Completing a launch opens the created resource’s detail page for status, output and lifecycle controls. These modes can incur compute charges separately from model usage. Creating a resource does not guarantee immediate capacity or successful execution.

For decision payloads, examples and billing details, see [Typed decisions](https://www.nodus-compute.ai/docs/guides/inference/decisions/).

The **Labs** filter identifies model authors. A `nodus/` model ID is an API alias, not a claim that Nodus trained the model.
