Bring your own model
Models → Add takes any Hugging Face repository laid out in WhisperKit's CoreML format — a fine-tune, a distilled variant, a model that did not exist when FreeFlow shipped — or a model folder already on disk. It downloads once, appears beside the six built-ins, and any mode can use it. Delete it from the same list.
What fits
FreeFlow runs Whisper models through WhisperKit, so any repository that ships WhisperKit's CoreML bundles works: the official argmaxinc/whisperkit-coreml conversions of every OpenAI Whisper size, distil-whisper variants, and fine-tunes that someone has converted. A plain PyTorch or safetensors checkpoint does not fit — it has to be converted to CoreML first, which WhisperKit's own tools do. Parakeet and Cohere Transcribe come through FluidAudio and are the built-in versions only.
- Find the repositoryBrowse Hugging Face for WhisperKit CoreML conversions. The repository's files should include a
MelSpectrogram.mlmodelc,AudioEncoder.mlmodelcandTextDecoder.mlmodelcset. - Models → Add → Hugging FacePaste the repository id (
owner/name) and, if the repo holds several variants, the variant folder name. FreeFlow downloads it into your data folder; the × beside the progress bar cancels, and a failed download says why. - Assign it to a modeThe model appears in every mode's model menu. Make a mode for it, or switch an existing one.
- DictateThe first take loads the model; afterwards it stays resident until you switch to a different engine.
From a folder
Models → Add → Folder points FreeFlow at a model directory already on disk — one you converted yourself, or copied from another Mac. It is imported into the data folder so the original can move.
Expectations
- Accuracy and speed are the model's, not FreeFlow's. A tiny fine-tune is fast and rough; a large one is slow and careful. The built-in Parakeet V3 is hard to beat for English and European languages.
- Languages come from the model. Whisper variants speak Whisper's 99; a fine-tune for one language may only do that one well.
- The dictionary applies to custom models too; formatting and modes work the same.
- Disk: a Large-class model is around 3 GB. Delete what you do not use from the Models list.
Questions people ask
Can I add a Parakeet or Cohere variant?
Not through Add. Those engines run through FluidAudio and ship as the built-in versions. Add is for WhisperKit-format models.
Does adding a model need the network?
Once, to download it. After that it runs offline like the built-ins.
Where does the model go?
Into FreeFlow's data folder (Settings → General → Data folder), with the built-ins. Deleting it from Models removes the files.
Will a model I add survive an update?
Yes. Models are files beside the app, not inside it.