Skip to content

Use the built-in web UI

Terminal window
mistralrs serve -m Qwen/Qwen3-4B --quant 4

When the server is ready, open http://localhost:1234/ui. Keep the terminal open while chatting; press Ctrl+C to stop. For installation, follow the quickstart.

The UI provides:

  • A chat panel with markdown rendering, code block syntax highlighting, and streamed output.
  • A settings drawer with sampling controls (temperature, top-p, top-k, max tokens) and a system prompt field.
  • Inline rendering of thinking tokens when the model emits them, with Default, Off, Low, Medium, High, and XHigh reasoning controls.
  • Tool-call visualization when --enable-search, --enable-code-execution, or --enable-shell is also set.
  • Tool approval cards when approval mode is ask.
  • Multimodal attachments when the loaded model supports images, audio, or video.
Terminal window
mistralrs serve --agent -m Qwen/Qwen3-4B

See tool calling for what --agent enables and the individual flags behind it.

When the model calls a tool, the UI renders a collapsed block in the conversation. Expanding shows the tool arguments and result. Search blocks display the query and a result count; code-execution blocks display the executed Python and any output or images; shell blocks display the commands, stdout/stderr, and exit status.

The settings drawer controls search, code execution, shell, and tool approval per conversation without restarting the server. Set Tool approval to ask to approve or deny each agent action inline.

Code and shell execution use the configured sandbox. See code execution and shell for setup.

Use the paperclip to attach images, audio, or video supported by the loaded model.

Modality support per model is in the supported models reference.

The settings drawer state is persistent per browser. System prompts and sampling parameters survive reloads.

Clearing browser local storage for the site resets all UI state.

Pass --no-ui to mistralrs serve to skip mounting the UI router. The HTTP API continues to serve normally.