Playground

The Playground tests models directly from the dashboard:

• Chat with request / response / timing / events / curl inspector tabs. When upstream llama.cpp timing metadata is present, the timing sidebar shows TTFT, prefill, decode, and stream-close.
• Speech (TTS) from plain text or a server-extracted article URL (paragraph-chunked), with waveform playback.
• Image and transcription testing.

Chat

The chat tab sends through llama-dash using a system API key generated for the dashboard playground. It supports:

• Model selection from local models and peers.
• System prompt and sampling controls.
• Streaming generation with stop, regenerate, edit, and fork actions.
• Per-response metrics such as TTFT, tokens per second, and output tokens.
• Inspector tabs for request, response, timing, events, and a reproducible curl command.

Shortcuts:

• Enter sends.
• Shift+Enter inserts a newline.
• r sends and re-runs when supported by the chat flow.

Speech

The speech tab targets /v1/audio/speech compatible models. It supports plain text input and article mode:

• Text mode speaks the textarea content.
• Article mode fetches a URL server-side, extracts readable text, lets you edit it, then generates segmented speech.
• Voice selection uses the model's reported voices when available, or accepts a manual voice name.
• Output is MP3 at 24 kHz with preview playback, segmented playback, and download actions.

Image

The image tab targets /v1/images/generations. Select a model, choose 256x256, 512x512, or 1024x1024, then submit a prompt. Generated images remain in the page history until cleared and can be downloaded.

Transcribe

The transcribe tab targets /v1/audio/transcriptions with verbose_json output. Drop or select an audio file, or record from the browser, then transcribe with the selected model. Supported upload types include MP3, OGG, WAV, FLAC, M4A, and WebM up to 20 MB.