Where Does Whisper Cache Models?
Where openai-whisper caches downloaded models on each OS, and how to clear them.
Last updated
openai-whisper downloads model files on first use into the user cache: ~/.cache/whisper on Linux and macOS, %USERPROFILE%/.cache/whisper on Windows. Each size is one .pt file plus a matching .pt hash sidecar.
The cache respects $XDG_CACHE_HOME when set, so relocation is one variable. GUI wrappers (MacWhisper, Buzz, MacWhisper-style apps) do not share this folder; they keep models under their own Application Support trees.
Where Whisper stores this, by platform
~/.cache/whisper/
One .pt per model size (tiny.pt through large-v3.pt). Respects $XDG_CACHE_HOME. GUI apps like MacWhisper use their own folders instead of this cache.
%USERPROFILE%\.cache\whisper\
Same .pt files under the user profile cache. Delete stale sizes to reclaim gigabytes; they re-download on demand.
~/.cache/whisper/
Same layout as Linux. Homebrew and pip installs share this cache; only GUI wrappers diverge.
Frequently asked questions
How do I clear or move Whisper models?
Delete the .pt files below to force re-download, or set XDG_CACHE_HOME to relocate the cache before first run. Keep at least one model you actually use; tiny/base download in seconds while large takes a while.
Do I need anything besides the cache?
No. The pip package plus ffmpeg handle transcription; models fetch on first use per size name (tiny through large-v3). MacWhisper and other GUIs keep their own model folders separately under Application Support.
Notice an outdated path? Let us know.