Playground
The Playground lets you talk to deployed models in the browser, to check their answers. Requests go through the console to the inference gateway; the gateway key stays on the server and never reaches the browser.
Chat
Open Playground → Chat:
- Pick a model at the top. The list is the deployments in Model Serving → Inference Services, ready ones first, shown as
release · served name; one that cannot be called yet is greyed out with why. If it says no model is deployed yet, deploy one; it appears here once ready. - Type below: Enter sends, Shift+Enter adds a line. Stop ends a generation early.
- Under every answer: time to first token (TTFT), duration, input/output tokens, tok/s and cache hit rate.
Working with messages:
- Edit or delete any message, or regenerate the last answer.
- Switch the composer to "Assistant": Add only appends the message, Send appends it and asks the model to continue. Useful for building few-shot examples.
- Reasoning models' thinking is shown in a collapsible block.
A conversation keeps one session ID, so the gateway pins it to one inference instance and the prefix cache keeps hitting. New chat starts a new session.
Parameters
Parameters opens the panel:
- System prompt: sent as the system message.
- Max output tokens, and under "More" temperature, top_p, seed, stop sequences and penalties.
- Empty parameters are not sent; the engine uses its own defaults.
- Reset restores the defaults and keeps the system prompt.
View code
View code gives the request for the current model, parameters and conversation as cURL, Python or Node.js. Set the MODELSPHERE_API_KEY environment variable to an API key first; see Calling the API.