AI Models
YapToText runs Whisper on your Mac to turn your voice into words, then runs a small language model on your Mac to clean those words up when the mode asks for it. Both models ship inside the app: Whisper Large v3 Turbo (Q5) for dictation and Phi-3.5 Mini Instruct for cleanup. There is nothing to download, no account, and nothing leaves the machine.
The AI Models page in the app is where you see those two models, swap either one for another from the library, or add a model file of your own. This article covers what each model does, what the library holds, and where the files live.
The two models that do the work
Two models run in sequence, and the AI Models page names them by their job. The dictation model turns your voice into words. That is Whisper Large v3 Turbo (Q5), a whisper.cpp model, and it ships inside the app. The cleanup model turns those words into finished text when the mode asks for it. That is Phi-3.5 Mini Instruct, a llama.cpp GGUF model, also inside the app. Both are MIT licensed.
Not every dictation touches the cleanup model. Raw Transcription types exactly what you said, with punctuation, and never touches the cleanup model. Clean Up, Note, Email, Message, and Code Comment each hand the transcript to the cleanup model with their own instructions. Every mode's editor has an AI cleanup toggle, so you can switch that stage off for any mode.
The speech model warms up when the app launches. The cleanup model loads the first time a mode needs it, which takes a moment, and the panel tells you while it happens. After that it is fast. A dictation's details in History show how long the dictation model took and how long cleanup took.
The model library
The page opens with Your defaults: the bundled dictation model and the bundled cleanup model, each under a plain label saying what job it does. A Language picker sits with the dictation model. It is the language hint the speech model receives.
Below that are two libraries. The Dictation Model Library holds the other Whisper models, from Whisper Tiny (English) at 75 MB up to the full Whisper Large v3 at about 3 GB, with Base, Small, Medium, and the other Large v3 Turbo variants in between. The Cleanup Model Library holds alternatives to the bundled cleanup model: Llama 3.2 1B and 3B Instruct, Qwen 2.5 1.5B and 3B Instruct, and Gemma 2 2B. Every entry shows its provider, its languages, and its download size.
Each row also carries a star rating. The stars open two plain ratings, one for accuracy and one for performance, plus a Recommended mark on the pick for most people. The rule of thumb is simple: bigger models hear better, smaller ones are lighter on battery and memory. Whisper Large v3 Turbo (Q5) is the recommended dictation model because it is both accurate and quick, and Phi-3.5 Mini Instruct is the recommended cleanup model.
A library model is downloaded once, over HTTPS from Hugging Face, with a progress bar you can cancel. Once installed, it can be selected as your default, shown in Finder, or removed. A model can only be selected when it is fully on disk. That download is the only network traffic the app ever makes, and it carries nothing about you or your dictations.
Bring your own model
Under Your own models you can add any whisper.cpp speech model (.bin) or llama.cpp cleanup model (.gguf) from disk. The app reads the first bytes of the file to tell which kind it is, so there is nothing to configure: a GGUF file becomes a cleanup model and a whisper GGML file becomes a dictation model. The file is copied into the app's Models folder, and the model then appears everywhere a model can be chosen, including the Energy page, where it is marked as yours.
A file that is neither format is rejected up front with a message saying so. A GGUF file can still be in an architecture the built-in engine does not support; the row's description warns that if it fails to load, that is why. Removing one of your own models removes its copied file too.
Models that follow the power source
The Energy page can run different models depending on how the Mac is powered. Turn on Switch models with the power source and you get a Plugged in and an On battery choice for the dictation model, and the same pair for the cleanup model. Every model you have appears in those menus, including ones you added yourself. Leaving a choice on Use the main selection keeps whatever the AI Models page says.
The cleanup model is the heaviest thing the app runs, so putting a lighter one on battery saves the most power. A mode can also carry its own model choice in its editor, and a mode's choice always wins over these defaults.
Where the models live
The Storage card says it plainly: models are stored locally on your Mac. Reveal in Finder opens the Models folder. Downloaded models and models you added live in the app's own data container, not inside the app itself, so they survive updates. The two bundled models ship inside the app and stay with it; they cannot be removed. Nothing in that folder is ever uploaded.
The code is public on GitHub, and the issues page is where to tell me if a model misbehaves. The privacy page covers what the app stores.
Questions
Does YapToText use Whisper for dictation?
Yes. It ships with Whisper Large v3 Turbo (Q5) running through whisper.cpp on your Mac, and the Dictation Model Library offers the other Whisper sizes if you want to trade accuracy for speed or the other way round.
Is the cleanup done by a local LLM?
Yes. The bundled cleanup model is Phi-3.5 Mini Instruct, a llama.cpp GGUF model that runs on your Mac. You can swap it for Llama 3.2, Qwen 2.5, or Gemma 2 from the library, or add any GGUF file of your own.
Can I use my own Whisper or GGUF model?
Yes. Add any whisper.cpp speech model (.bin) or llama.cpp cleanup model (.gguf) under Your own models. The app detects which kind it is from the file itself and lists it wherever a model can be chosen.
How much disk space do the models take?
The two bundled models come to roughly 3 GB as the app lists them: 574 MB for Whisper Large v3 Turbo (Q5) and 2.3 GB for Phi-3.5 Mini Instruct. Every library entry shows its download size before you fetch it, and any download can be removed later.
Does the app need the internet to run the models?
No. Both bundled models work offline from the first launch. The only time the app touches the network is when you start a download from the model library, and that request carries nothing about you or your dictations.