Where apps hide your files

Finding your Whisper model files

Updated 2026-10-28 · 6 min read

Where Whisper keeps its models depends on the tool: OpenAI's Python package, whisper.cpp, MacWhisper and the Hugging Face based tools each use their own folder. Here is each path, how to check sizes, and what downloads again if you delete it.

The locations

Each Whisper tool keeps its models in its own place. The four you are most likely to have on a Mac:

~/.cache/whisper/                                 (OpenAI's whisper Python package)
<your whisper.cpp folder>/models/                 (whisper.cpp)
~/Library/Application Support/MacWhisper/         (MacWhisper)
~/.cache/huggingface/hub/                         (faster-whisper, mlx-whisper)

~/.cache starts with a dot and ~/Library is hidden, so Finder shows neither by default. In Finder choose Go > Go to Folder (⇧⌘G), paste the path and press Return, or press ⇧⌘. in a Finder window to show hidden items. From Terminal, open ~/.cache/whisper opens the folder if it exists.

OpenAI's whisper package

The openai-whisper package downloads a model the first time you ask for it by name and keeps it in ~/.cache/whisper. In its source, the folder is $XDG_CACHE_HOME/whisper when that variable is set, and ~/.cache/whisper otherwise. Each model is one PyTorch file named after the model, such as base.pt, small.en.pt or medium.pt. Two names are aliases: asking for large saves large-v3.pt, and asking for turbo saves large-v3-turbo.pt, so there is no file called turbo.pt to look for.

To see what you have and how much space it takes, without changing anything:

ls -lh ~/.cache/whisper
du -sh ~/.cache/whisper

No folder means the package has never downloaded a model for your user account, or you always point it elsewhere: the whisper command takes --model_dir, and whisper.load_model() takes download_root.

whisper.cpp

whisper.cpp has no fixed location. Models live in the models folder of wherever you cloned or unpacked it, as ggml-<name>.bin files. The project's download script fetches them:

sh ./models/download-ggml-model.sh base.en

That saves models/ggml-base.en.bin. According to the script's source, it saves next to itself by default, saves into the current folder instead when the script sits in a folder named bin, and accepts a target folder as a second argument. The name tells you the variant:

Part of the name	Meaning
tiny, base, small, medium, large-v3, large-v3-turbo and so on	Model size
.en	English-only model
-q5_0, -q5_1, -q8_0	Quantized, smaller file
-tdrz	Tinydiarize speaker-turn model, fetched from a separate repository

If you built Core ML support, the README's generation script also creates a folder such as models/ggml-base.en-encoder.mlmodelc beside the .bin file. It is a folder, not a file, so copy or delete it as a whole.

Apps that bundle whisper.cpp, and package-manager installs, can keep models somewhere else. Check that app's own settings first.

MacWhisper

MacWhisper's help page gives its data folder as:

~/Library/Application Support/MacWhisper

The same page says this folder holds Models (the downloaded speech-to-text models), Database (transcripts, sessions and metadata) and External Media. The sandboxed edition, which the page calls Whisper Transcription, keeps the same data inside its container:

~/Library/Containers/com.goodsnooze.MacWhisper/Data/Library/Application Support/MacWhisper

Usually only one of the two exists. The help page lists preferences at ~/Library/Preferences/com.goodsnooze.MacWhisper.plist and regenerable data at ~/Library/Caches/com.goodsnooze.MacWhisper. Because your transcripts sit beside the models, never delete the whole folder to free space. The vendor says the data can be removed by hand, but not what removing a model file does to the app, so remove models from MacWhisper's own settings and use the folder only to check sizes:

du -sh ~/Library/Application\ Support/MacWhisper/* 2>/dev/null

faster-whisper and mlx-whisper

These two fetch models from Hugging Face by repository name. faster-whisper maps base, large-v3 and the other sizes to Systran/faster-whisper-* repositories (its turbo comes from a mobiuslabsgmbh repository) and downloads them with huggingface_hub.snapshot_download, the same function mlx-whisper uses. Unless you pass a folder (faster-whisper's download_root), the files land in the Hugging Face cache, which the huggingface_hub source puts at ~/.cache/huggingface/hub. The HF_HOME and HF_HUB_CACHE variables move it.

Each model is a folder named like models--Systran--faster-whisper-base. To list the Whisper ones with their sizes:

du -sh ~/.cache/huggingface/hub/models--*whisper* 2>/dev/null

These hold CTranslate2 or MLX files, so a whisper.cpp ggml-*.bin file cannot stand in for one, and the reverse is also true.

Finding stray copies

If you cloned whisper.cpp more than once, or cannot remember where, this lists every ggml-*.bin file up to four folders deep in your home folder, skipping ~/Library. It only reads names:

find ~ -maxdepth 4 -path ~/Library -prune -o -name "ggml-*.bin" -print 2>/dev/null

macOS may ask whether Terminal can access Documents or Desktop while it runs. Saying no only skips those folders.

Freeing space, and what comes back

Quit the tool first, then move unwanted files to the Trash in Finder:

• OpenAI's package: a deleted .pt file is downloaded again the next time you load that model, and the source also re-downloads a file whose checksum does not match. Nothing else is kept in that folder.
• whisper.cpp: delete the ggml-*.bin files (and their -encoder.mlmodelc folders) you no longer use, and leave the scripts. Running the download script again brings a model back.
• Hugging Face cache: other tools share it, so remove a single models--...whisper... folder, never the whole hub folder. A removed model downloads again on next use.
• MacWhisper: use the app, as above.

The large models are big downloads, so delete only what you will not need soon.

To keep models on an external drive, point each tool at it: --model_dir or download_root for OpenAI's package, a folder you choose for whisper.cpp (the script's second argument, then -m with the full path), and HF_HUB_CACHE for the Hugging Face tools. Setting XDG_CACHE_HOME also works for OpenAI's package, but it moves the Hugging Face cache and other tools' caches with it.

Why Spotlight comes up short

Spotlight and Finder searches leave out ~/Library, app containers and hidden dot-folders such as ~/.cache, which is where three of these four tools keep their files. A whisper.cpp clone in an ordinary folder can turn up, but only when you remember the exact file name.

Finding a model next time

In Everywhere, type ggml- to list every whisper.cpp model with its folder, or large-v3 to see each tool's copy of that model, including the files and folders inside ~/.cache. ⌘↵ reveals one in Finder. The sandboxed MacWhisper edition keeps its data in ~/Library/Containers, which Everywhere covers once it has Full Disk Access, and the search matches names only, not what is inside a model.

Common questions

Where does OpenAI Whisper store models on a Mac?

In ~/.cache/whisper, or $XDG_CACHE_HOME/whisper if that variable is set. Each model is one file such as base.pt; the turbo model is saved as large-v3-turbo.pt.

Where are whisper.cpp models stored?

In the models folder of the whisper.cpp folder you cloned, as ggml-<name>.bin files. The download script also accepts a different target folder as its second argument.

Where does MacWhisper keep its models?

In ~/Library/Application Support/MacWhisper, next to its transcript database. The sandboxed Whisper Transcription edition keeps the same folder inside ~/Library/Containers/com.goodsnooze.MacWhisper.

Is it safe to delete ~/.cache/whisper?

It holds only the OpenAI package's model files, so deleting them frees the space and each model downloads again the next time you load it. MacWhisper and the Hugging Face tools keep their models elsewhere.