Clean-up
A small language model on your computer tidies what you said before it is typed. Choose Off, Light or Polished, and a different mode for particular apps.
Speech comes out as speech: "um", false starts, no full stops. Clean-up runs a small language model on your computer to tidy the transcript before it is typed. Nothing leaves your computer. Settings are under Clean-up.
Modes
| Mode | What you get |
|---|---|
| Off | Exactly what the speech model heard, trimmed. |
| Light (default) | Punctuation, capitals and paragraphs; fillers, stutters and false starts removed; "no wait", "I mean", "actually" and "scratch that" applied, keeping only what you meant; "new line" and "new paragraph" become line breaks. Your wording stays, and names, numbers and code are kept exactly as spoken. |
| Polished | Light, plus light grammar fixes and a tone that suits the app: short and casual in chat, courteous in email, terse in a terminal or code editor (no curly quotes), prose in documents, precise in an AI prompt. |
Both modes follow your language, spelling and dictionary.
If clean-up ever cannot run (the model is still loading, or is missing), your words are typed as heard. A dictation is never lost because of clean-up.
Choose a model
Model picks which language model does the work; Model status downloads it. Qwen3 4B Instruct is the default and gives the best results. Gemma 3 1B is much smaller and faster, and fine for Light. See Models.
Until a clean-up model is downloaded, clean-up stays off and dictation keeps working.
Per-app rules
Use a different mode in particular apps: Off in a terminal keeps commands and identifiers exactly as spoken, Polished suits email.
Under Per-app rules, pick the App and the Mode, then Add rule. Only apps with a window open right now are listed, so open the app first.
On Linux, per-app rules need MojoFlow to see which app is in front. On GNOME that is the helper extension from installation (sign out and back in once after installing). If MojoFlow cannot tell, the Clean-up page says why, and the rules apply everywhere.
Share AI Workbench's model
If you use AI Workbench, MojoFlow can use the language model the workbench already has running instead of loading its own, so your computer holds one model instead of two. Turn on Use the workbench's model under Mojo Up AI Workbench. The Clean-up page says which model is doing the work, and MojoFlow goes back to its own by itself whenever the workbench's is not available.