Use the Playground

Try a model in the Playground before you write any code: send it input and see its reply in TIR. The layout depends on the model’s task: a text model is a chat window, a speech model takes audio, and an image model shows a picture back.

Before you begin

All you need is a TIR project.

Find a model to try

1. In TIR, open your project. In the side menu’s PRODUCTS list, select GenAI API. The Models page lists every model you can use, one per row.
2. Use the search box to filter by name or author, or the Author and Type filters to narrow the list.
3. Select a model’s row to open its own page, with Playground, Usage, Get Code, and, for Text-to-Text models, Async History. To move to another model, use the switcher on the page.

A model that shows Request instead of opening is out of inventory. Select Request to ask for it, and TIR emails you once it is available. A model’s own page also offers Benchmarking Request, which asks E2E to benchmark it for your use case. Both place a support ticket.

Send a prompt to a text model

A text model’s Playground is a chat window: type a prompt in the box and send it. There is no separate system-prompt field. Turn on Stream to see the reply as it is generated instead of waiting for the whole answer.

Set Advanced Parameters

The Advanced Parameters panel is a collapsible section next to the chat. Each model sets which parameters it shows, with their own minimum, maximum, and default. For the e5-mistral-7b-instruct model today, the panel offers:

Parameter	Range	Default
Top P	0 to 1	1.0
Temperature	0 to 1	0.5
Max Tokens	1 to 300	300
Presence Penalty	-2.0 to 2.0	0
Frequency Penalty	-2.0 to 2.0	0

Other models have their own values, including a different Max Tokens limit.

Top P limits the model to its most likely next words: a lower value gives a more focused reply, a higher value a more varied one. Temperature controls how random a reply is: values near 0 are more predictable, values near 1 are more creative. Max Tokens sets how long a reply can be. Presence Penalty discourages the model from returning to a topic it already covered, and Frequency Penalty discourages it from repeating the same words or phrases. A higher value on either means less repetition.

Max Tokens applies to one reply and is separate from the model’s context length. See My reply stops partway through for how a model’s context length works.

An embedding model and the moderation model, Llama Guard, have no Advanced Parameters panel.

Try a speech-to-text model

A speech-to-text model’s Playground takes audio instead of text: record it with your microphone, or upload a .wav, .flac, or .mp3 file. A recording you make in the Playground is captured at 16 kHz mono and sent as a WAV file. An uploaded file can be up to about 49.6 MB for Whisper Large V3, and up to 10 MB for the other speech-to-text models.

Whisper Large V3’s Playground also lets you choose Task (Transcribe or Translate), a Language, and whether the output carries a Return Timestamp per word, per sentence, or not at all. One speech-to-text model transcribes live from the microphone instead of taking a file. As you speak, it shows a running transcript, a word count, and a Recording indicator, and you can copy the transcript or download it as a .txt file.

Try a text-to-speech model

A text-to-speech model that clones a voice needs a reference audio sample and its matching reference text, as well as the text you want spoken. A simpler text-to-speech model instead needs only the text, a Speaker (male or female), and a Language.

Try an embedding model

An embedding model’s Playground looks like the chat box, but its reply is a vector, a list of numbers. The Playground shows it shortened, with a Copy Embeddings action.

Try an image-generation model

An image-generation model’s reply is shown as an image, with Fullscreen and Download Image actions.

Reset the conversation

A Playground conversation lasts only while the page is open. Select Reset to clear it and start again.

Read the result

A successful reply comes back in the shape its model type uses: text in the chat, a vector for an embedding model, an image for an image model, or a transcript for a speech model.

A failed reply in a chat model shows in the chat, marked as an error, with the server’s own message. A non-chat model’s failure shows as a pop-up message instead. If you try to transcribe before you add a file, or TIR cannot reach your microphone, a message tells you what to fix.

To send many prompts at once instead of one at a time, see Run bulk prompts. To call the same model from your own code, see Call the API with the OpenAI SDK, Call the API with REST, or Call the API with the TIR SDK. For help with any message, see Troubleshooting TIR GenAI API.