Petal is a native macOS app for fast, local-first audio transcription in a clean, minimal interface.
petal-transcibe-file-to-clipboard.mp4
Petal W1 is Petal's on-device cleanup model, fine-tuned from Qwen3.5-2B. It turns a rambling dictation into the short message you meant, in your own voice, in about 0.2 s. Claude Sonnet graded each output on 120 held-out real dictations.
| Model | Error-free outputs | Meaning errors per dictation | Length of long rambles | Answered the dictation |
|---|---|---|---|---|
| Petal W1 v1.4 | 47% | 0.56 | 55% | 0 |
| Petal W1 v1.2 | 40% | 0.73 | 80% | 0 |
Petal can also clean up your dictation with a cloud model. Use your own OpenAI, Anthropic, or OpenRouter key, or any OpenAI-compatible server such as Ollama or LM Studio. Open Settings > Intelligence and choose Cloud Model.
- Start from a preset (Clean Up, Email, Notes, Professional, AI Prompt, or Assistant) or write your own system prompt, then test it in Try It.
- Insert variables that Petal fills in on your Mac each time:
{{name}},{{first_name}},{{app}},{{window}},{{language}},{{region}}, and{{time_zone}}. - Turn on Date and Time to turn words like "next Friday" into exact dates. Turn on Web Search to let the model look up facts that you ask for.
- Turn on Screen to send a screenshot of your screen with each dictation, so the model can spell the names it sees and know what "this" means. Screen needs Screen Recording access and is not available for custom servers.
- In Settings > Router, each app or website route can use its own intelligence (Apple Intelligence, Petal W1, or Cloud Model) instead of the default. A Cloud Model route uses the provider, model, and tools from Settings > Intelligence.
- Petal keeps your keys in the macOS keychain and sends the transcript, and the screenshot when Screen is on, only to the provider that you choose.
