• AI tools
  • Developer

Transcription apps download speech models of up to 3 GB, sometimes more than once

Whisper-based apps keep models from 75 MB to over 3 GB. Where MacWhisper, whisper.cpp and Python tools store them, and how to keep only one copy.

4 min read

Local transcription is excellent on Apple silicon. It depends on a speech model, and the most accurate ones are large. Try two apps and a command-line tool, and the same model can be on disk three times.

Model sizes

ModelApproximate size
tinyAbout 75 MB
baseAbout 140 MB
smallAbout 470 MB
mediumAbout 1.5 GB
largeAbout 3 GB

Figures are for the common whisper.cpp model files. Other formats vary.

Where they are

du -sh ~/.cache/whisper 2>/dev/null                          # Python whisper
du -sh ~/.cache/huggingface 2>/dev/null                      # Hugging Face tools
find ~/Library ~ -maxdepth 4 -iname '*ggml*.bin' -exec du -h {} + 2>/dev/null   # whisper.cpp files
find ~/Library/Application\ Support ~/Library/Containers -maxdepth 1 -iname '*whisper*' -exec du -sh {} + 2>/dev/null

Keep one copy

  • Delete models you tried and did not use. Apps download them again if needed.
  • Most people need one model. small or medium is a common balance of speed and accuracy.
  • Some apps let you choose the model folder. Pointing them at one folder avoids duplicates.

Related: local LLM models and machine learning caches.

Common questions

How big are Whisper models?

From about 75 MB for tiny to about 3 GB for large, in the common whisper.cpp format.

Where does Whisper store models on a Mac?

Python whisper uses ~/.cache/whisper, Hugging Face tools use ~/.cache/huggingface, and apps usually keep them in their own Application Support or container folder.

Can I delete Whisper models?

Yes. They are downloads and are fetched again when an app needs them.

Read next