Source: https://fnflow.app/mlx-whisper-vs-whisper-cpp

# mlx-whisper vs whisper.cpp
Both run Whisper on an Apple Silicon Mac. mlx-whisper is the easy choice if you're in Python and already use MLX. whisper.cpp suits you better if you want a self-contained binary to ship in an app or run without Python. We built FnFlow on whisper.cpp, and we haven't benchmarked mlx-whisper against it, so you won't find a speed verdict here. You will find the commands to get your own.

## At a glance

|  | mlx-whisper | whisper.cpp |
|---|---|---|
| Runs on | Apple Silicon, through MLX | Apple Silicon through Metal; also Intel, Linux, Windows and more |
| Install | pip install mlx-whisper and ffmpeg | Homebrew or CMake |
| Use from | Python or its CLI | CLI, server, C API |
| Model format | MLX, from Hugging Face or converted | ggml |
| Default model | tiny | You choose the file |
| Quantization | 4-bit through the convert script | q5_0, q5_1, q8_0 and others |
| Word timestamps | Yes | Yes |

From github.com/ml-explore/mlx-examples (whisper) and github.com/ggml-org/whisper.cpp, checked 6 November 2026.

## Pick mlx-whisper if

- You're writing Python scripts or notebooks on a Mac.
- Your project already uses MLX.
- You want a model from Hugging Face by name and a two-line start.

## Pick whisper.cpp if

- You're shipping an app and don't want to bundle Python. This was our reason.
- You want to keep a model loaded behind a small HTTP server, with whisper-server.
- The same code has to run on other systems.

## Measure it yourself

Use 16 kHz mono WAV and the same model in both. Run each one twice and keep the second time, because the first includes downloads and shader compilation.

time mlx_whisper test.wav --model mlx-community/whisper-large-v3-turbo

time whisper-cli -m ggml-large-v3-turbo.bin -f test.wav -l auto
Check the text as well as the clock. Compare with the same language setting, and read the result on your hardest sentences. [Our whisper.cpp benchmark](whisper-cpp-apple-silicon-benchmark) describes the word-diff method we used, which you can apply to this comparison. It found almost no change from f16 to q8_0 and a real change at q5_0, so match quantization levels before comparing speed.

## If you're here to dictate

Neither library will type into Cursor for you. FnFlow handles that part: hold fn, talk, release, and whisper.cpp's text is pasted where your cursor is. [Which Whisper model to use](which-whisper-model-mac) covers the model choice.

## Questions

**Is mlx-whisper faster than whisper.cpp?**
We haven't measured it, and results depend on the chip, model and quantization. Use the commands on this page with your own audio.

**Does mlx-whisper need Python?**
Yes. You install it with pip, and it also needs ffmpeg. whisper.cpp needs neither for WAV input.

**Which models work?**
mlx-whisper loads MLX-format models from Hugging Face. The mlx-community organization has converted ones, including large-v3-turbo. The default is tiny. whisper.cpp uses ggml files.

**Can I quantize with both?**
Yes. mlx-whisper's convert script makes 4-bit models with -q. whisper.cpp has a quantize tool with q5 and q8 methods, and we measured those in our benchmark.
