Source: https://fnflow.app/run-whisper-locally-mac

# How to run Whisper locally on an Apple Silicon Mac
Whisper runs well on Apple Silicon without Python or a cloud account. We use whisper.cpp inside FnFlow, so these are the steps and numbers from our own setup: installing it, choosing a model, preparing audio and keeping the model loaded between requests.

## Install it

The quickest route is Homebrew. The formula is now called whisper.cpp (it used to be whisper-cpp) and was at version 1.9.5 when we checked on 10 October 2026.

brew install whisper.cpp
whisper-cli -h
It installs whisper-cli but no models. The current formula also builds with the server example turned off, so you won't get whisper-server from it. For that, build from source:

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
cmake -B build
cmake --build build -j --config Release
ls build/bin
You don't need any extra flags for the GPU. On Apple platforms ggml's CMake turns Metal on by default and embeds the Metal library in the binary. If you plan to copy the binaries to other Macs, don't tune them for your own chip. We build FnFlow's copy for armv8.4-a+fp16 rather than native, because a build tuned for an M4 Pro wouldn't start on an M1.

## Download a model

From the source checkout, the bundled script fetches converted ggml models from Hugging Face into models/:

sh ./models/download-ggml-model.sh large-v3-turbo-q8_0
sh ./models/download-ggml-model.sh small-q5_1
With Homebrew, download the file directly. The URL pattern is the same one the script uses:

curl -L -o ggml-large-v3-turbo-q8_0.bin \
https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q8_0.bin

## Which model to pick

On 8 October 2026 we ran 32 real dictations through four models on an M4 Pro. Every timing below is a cold whisper-cli run that includes loading the model. The audio was mostly Russian with English technical terms, and the full f16 model was the reference.

| Model | Download | Words different from the best model | Same text, word for word | Time per dictation |
|---|---|---|---|---|
| large-v3-turbo (full, f16) | 1,625 MB | reference | – | 2.14 s |
| large-v3-turbo q8_0 | 874 MB | 0.0% | 31 of 32 | 1.98 s |
| large-v3-turbo q5_0 | 574 MB | 2.4% | 21 of 32 | 2.21 s |
| small q5_1 | 190 MB | 22.0% | 0 of 32 | 0.84 s |

Our pick is large-v3-turbo q8_0. It's half the size of f16 and gave the same text in 31 of 32 files, and it was slightly faster. q5_0 saves another 300 MB but changed words in 11 files. small q5_1 is fast and fine for a first test, but it got a fifth of the words different from the reference. Tiny and base turned Russian into nonsense in an earlier run. The full write-up is in our [whisper.cpp benchmark](whisper-cpp-apple-silicon-benchmark).

## Convert audio to 16 kHz WAV

whisper-cli reads 16-bit WAV, so convert voice memos and MP3s first. macOS ships afconvert, and ffmpeg works too:

afconvert -f WAVE -d LEI16@16000 -c 1 memo.m4a memo.wav
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav

## Transcribe a file

./build/bin/whisper-cli -m models/ggml-large-v3-turbo-q8_0.bin -f memo.wav -l auto -nt
-l auto detects the language, since the default is English. -nt drops timestamps, and -otxt writes a .txt file next to the audio. If names keep coming out wrong, pass them as a hint with --prompt "Kubernetes, Postgres, pgvector".

The first run of a fresh build is slow while Metal compiles its pipelines. On our M4 Pro that took 15 to 17 seconds, and later runs skip it.

## Keep the model warm

Every whisper-cli call reads the model from disk again. whisper-server loads it once and answers HTTP requests:

./build/bin/whisper-server -m models/ggml-large-v3-turbo-q8_0.bin -l auto -fa --host 127.0.0.1 --port 8080

curl 127.0.0.1:8080/inference \
-F file=@memo.wav \
-F response_format=json
For a 7.6-second sentence this cut our time from 2.26 s to 1.17–1.19 s, with identical text. Keep the server on 127.0.0.1. The whisper.cpp README warns against running it with admin rights or exposing it outside a sandbox, because it accepts file uploads. The --convert flag lets it accept other formats but requires ffmpeg.

whisper.cpp can also run the encoder on the Neural Engine through Core ML (-DWHISPER_COREML=1 plus a model generated with Python tools). We haven't measured that path.

## Or use FnFlow if you just want dictation

If the goal is talking instead of typing, FnFlow does all of the above for you. It has whisper.cpp built in and ships small q5_1 inside the installer, so it works offline right after install. The q8_0 model downloads in the background on Wi-Fi. A local server starts when you press fn and loads the model while you talk. When you let go, the text is pasted wherever your cursor is.

Other runtimes: [whisper.cpp vs faster-whisper](whisper-cpp-vs-faster-whisper-mac) and [mlx-whisper vs whisper.cpp](mlx-whisper-vs-whisper-cpp).

## Questions

**Do I need Python or PyTorch?**
Not for whisper.cpp. You only need a C++ toolchain and CMake to build it, or Homebrew to install it. Python comes in only if you want to generate a Core ML encoder for the Neural Engine.

**Why does my non-English audio come out wrong?**
whisper-cli defaults to -l en. Pass the language code, for example -l de, or -l auto to let Whisper detect it.

**Is there an English-only model?**
The download script has .en variants for tiny, base, small and medium. There's no .en version of large-v3-turbo. We haven't benchmarked the .en models.

**Can I use Homebrew's whisper-server?**
Not with the current formula. It builds whisper.cpp with the server example switched off, so build from source if you need whisper-server.
