How do you download Whisper and run it locally?
OpenAI doesn't offer a Whisper app or installer. To run Whisper on your own computer, you install Python and a free tool called ffmpeg, then add Whisper with one Python command. It downloads a model the first time you use it. We did this on every Python version from 3.9 to 3.15, and all seven worked. So the version isn't the hard part. Watch for two other things: without ffmpeg, Whisper skips your file and finishes as if it worked, and uninstalling it with pip leaves nearly a gigabyte behind.
Looking for a Whisper app in an app store? Several unrelated apps share the name, and one was an anonymous social app. This page is about OpenAI's speech-to-text model. Only want the transcript, with nothing to install? Skip to that.
What do you get when you download Whisper?
Two things: a set of Python packages, and a model file. The packages are Whisper itself plus PyTorch, the machine-learning library it runs on, and PyTorch is most of the size. On our test Mac, the standard install took 953 MB and 19 seconds before a single model was downloaded.
The model comes next, the first time you transcribe something. It ranges from 76 MB for tiny to 3.1 GB for large-v3. Turbo, the default, is 1.6 GB, so a working Whisper with its default model takes about 2.6 GB.
Disk space for a working Whisper
Which model to pick depends on your computer and your audio. Our test of ten Whisper models shows how many errors each one made, and which to pick.
Which Python version does Whisper need?
Any current one. Whisper's README says it's expected to work with Python 3.8 to 3.11. That's out of date. We gave the latest release, openai-whisper 20250625, a fresh environment on seven versions, then transcribed the same short clip with each one.
| Python | Installed | Transcribed | PyTorch |
|---|---|---|---|
| 3.9 | Yes | Yes | 2.8.0 |
| 3.10 | Yes | Yes | 2.14.1 |
| 3.11 | Yes | Yes | 2.15.0 |
| 3.12 | Yes | Yes | 2.15.0 |
| 3.13 | Yes | Yes | 2.15.0 |
| 3.14 | Yes | Yes | 2.15.0 |
| 3.15 | Yes | Yes | 2.15.0 |
WhisperAI's test, November 2, 2026: Apple M5 Pro, macOS 27, tiny model, a 5-second English clip. We didn't test 3.8.
If you're installing Python just for this, take 3.11 or later: 3.9 and 3.10 got older PyTorch releases.
How do you install Whisper on Windows?
Install Python from python.org. Its installer adds the python and py commands. Then install ffmpeg with a package manager (Chocolatey or Scoop), make a separate environment for Whisper, and install it there:
choco install ffmpeg
python -m venv whisper-env
whisper-env\Scripts\activate
python -m pip install -U openai-whisper
whisper interview.mp3 --model turboscoop install ffmpeg.The last line transcribes a file and saves the transcript as text and subtitle files in the folder you ran it from.
If Whisper prints "Skipping interview.mp3 due to FileNotFoundError" and no transcript appears, ffmpeg isn't installed or Windows can't find it. Install ffmpeg, then run the command again.
How do you install Whisper on a Mac?
Use Homebrew, the Mac package manager, to install ffmpeg. Then install Whisper inside its own environment, a folder that keeps its packages apart, because Homebrew's Python refuses a plain pip install: in our test it stopped with "externally-managed-environment".
brew install ffmpeg
python3 -m venv whisper-env
source whisper-env/bin/activate
python -m pip install -U openai-whisper
whisper interview.mp3 --model turboOn Linux, the steps are the same after sudo apt update && sudo apt install ffmpeg (Ubuntu or Debian) or sudo pacman -S ffmpeg (Arch). If the install fails with "No module named 'setuptools_rust'", run pip install setuptools-rust and try again.
Where does Whisper download its models?
To a folder in your home folder: ~/.cache/whisper, or C:\Users\you\.cache\whisper on Windows. Each model you run adds one file there, and Whisper reuses it next time.
Your environment folder
whisper-env
Whisper, PyTorch and 21 other packages
953 MB
The model cache
~/.cache/whisper
One file per model you've run
76 MB to 3.1 GB each
ffmpeg
Homebrew, Chocolatey or Scoop
Reads your audio and video
Stays when you delete the other two
Getting the space back takes more than an uninstall. Running pip uninstall openai-whisper in our test freed just 2 MB of the 953 MB, because PyTorch and the other packages stay. To remove Whisper completely, delete the whisper-env folder and the .cache/whisper folder.
Is there a Whisper desktop app?
Not from OpenAI. Other developers have built apps that run Whisper on your computer with buttons instead of commands. Our guide to Whisper desktop apps compares them. For a lighter install, whisper.cpp is a C/C++ port of Whisper that skips Python, though its quick start has you build it from source code.
Can you use Whisper without downloading it?
Yes. There are ways to use Whisper online in a browser, each with its own catch.
And if what you're after is the transcript, not Whisper, you can skip every step above. WhisperAI is a transcription app in your browser, with nothing to install and no model to pick. Upload a recording up to 5 GB and 10 hours, turn on speaker labels, and export it as DOCX, PDF, TXT or SRT. It doesn't run Whisper and isn't made by OpenAI. The Free plan covers 5 minutes a month with no card, and Business Pro is unlimited for $24.99 a month.
Frequently asked questions
Is Whisper free?
Yes. OpenAI released Whisper's code and model weights under the MIT License, so downloading and running it costs nothing but your disk space and time.
Does Whisper need a GPU?
No. Our test ran on a Mac's processor. A graphics card makes it faster: OpenAI's README lists about 1 GB of graphics memory for tiny and about 6 GB for turbo.
Does WhisperAI have an app?
Yes, for iPhone. WhisperAI also works in your browser, and paid plans add a Chrome extension.