Skip to content

Playground

The Playground is an interactive chat where you send prompts to a model that is currently serving and watch it stream a reply live. Use it to sanity-check a model or a fine-tune before you wire it into an agent.

Open it from the Develop sidebar, or land directly on a model with /studio/playground?modelId=<model> (the Models page deep-links here).

The model picker lists only models that are serving — models whose status is running, plus proxy (BYO-LLM) models. Anything not currently served is hidden.

If nothing is serving, the page shows an empty state (“Start a model to use the playground.”) with an OPEN MODELS → button that takes you to /studio/models. Start a model there first, then come back.

At the top of the workspace a dropdown shows the selected model (with a green “serving” dot). Open it to switch to any other serving model.

  • If you arrived via ?modelId=, that model is pre-selected when it is serving.
  • When the Playground is embedded elsewhere with a locked model, the picker is read-only and shows just the model name.
  • Switching models starts a fresh chat rather than reloading the previous model’s conversation.

The center pane is a chat thread. Type a prompt and the model streams its response token by token. Conversations are held per model, so each model you pick has its own thread history.

The right-hand Parameters panel controls sampling. Each is a slider:

Parameter Range Step
Temperature 0 – 2 0.05
Top P 0 – 1 0.05
Top K 1 – 100 1
Max Tokens 64 – 16384 64
Repetition Penalty 1.0 – 2.0 0.05

Below the sliders is a row of Presets. Clicking one applies a bundle of sampling settings (and a matching system prompt) in one go. The active preset is highlighted, and each button shows its temperature (t=…).

Preset Temp Style
Default 0.7 General helpful assistant
Creative 1.0 Imaginative, vivid writing
Precise 0.1 Factual, concise, verified-only
Code 0.2 Clean, documented code with step-by-step reasoning
CX Agent 0.3 Empathetic customer-support persona

The Playground talks to the same served endpoint your agents would use — it does not deploy or change anything. It is a scratchpad for trying prompts and tuning sampling parameters against a live model. To make a model available here, deploy/serve it from Models.