Technical explainer · Verified 2026-09-05

whisper.cpp
GGML and local inference

whisper.cpp is a C/C++ implementation for running Whisper speech recognition locally. It loads converted GGML model files, supports integer quantization, and offers acceleration paths including Metal and Core ML on Apple silicon. Its runtime and file format are separate from the Whisper checkpoint name, so the runtime documentation gives the relevant disk size and memory requirements.

whisper.cpp GGML disk sizes
ModelDisk
tiny / tiny.en75 MiB
base / base.en142 MiB
small / small.en466 MiB
medium / medium.en1.5 GiB
large-v12.9 GiB
large-v22.9 GiB
large-v2-q5_01.1 GiB
large-v32.9 GiB
large-v3-q5_01.1 GiB
large-v3-turbo1.5 GiB
large-v3-turbo-q5_0547 MiB

Disk sizes from the models README, in its original units. These are converted GGML files, not runtime RAM requirements or original checkpoint download sizes.

What a GGML model contains

whisper.cpp converts the original models into a custom binary format that packs model parameters, Mel filters, vocabulary and weights into a single file. The project describes a plain C/C++ implementation and lists platforms including macOS, Windows, Linux, iOS and Android. A GGML filename identifies a converted model for that runtime; it is not the original training checkpoint's file format.

Quantization reduces storage and memory

The project supports integer quantization and states that quantized models require less memory and disk space. Its models table includes q5_0 variants beside the larger converted files. Keep disk requirements separate from inference memory: the main README reports ~273 MB memory for tiny and ~3.9 GB for large. The project publishes no turbo memory figure in that runtime table.

Apple acceleration and licensing

On Apple silicon, whisper.cpp supports Metal and Core ML alongside ARM NEON and the Accelerate framework. The encoder can execute on the Apple Neural Engine through Core ML. Those are runtime acceleration options, not separate Whisper model families. The runtime is MIT licensed. When selecting a build, consult its platform instructions and keep the checkpoint, conversion and acceleration configuration explicit.

How WhisperJot uses this

WhisperJot's Windows and Linux Jot Local Pro uses whisper.cpp-format GGML models through a .NET runtime. Available choices include tiny, base and small, their q5_1 variants, large-v3-turbo at approximately 1.62 GB, and large-v3-turbo-q5_0 at approximately 574 MB. These product download estimates use different units from the upstream GGML table. On macOS, Jot Local Pro instead uses WhisperKit. Local transcription works offline after the one-time model download.

Explore WhisperJot →

Frequently asked questions

Does whisper.cpp run without a GPU?

Yes. The project provides a plain C/C++ implementation and CPU execution, with optional acceleration paths. Its platform support includes Windows, Linux and macOS. The model you choose still determines how much storage and memory you need. Consult the runtime's own memory table.

How large is the quantized turbo GGML file?

The upstream models table lists large-v3-turbo-q5_0 at 547 MiB. WhisperJot's Windows and Linux download estimate is approximately 574 MB. Preserve those original units when comparing the two listings; neither is a claim about turbo's runtime RAM use. Disk and memory are separate specifications.

Does whisper.cpp use Core ML on Mac?

It can execute the encoder on the Apple Neural Engine through Core ML, and the project also supports Metal. WhisperJot's macOS Jot Local Pro uses WhisperKit instead. Its Windows and Linux Jot Local Pro uses whisper.cpp-format GGML models. Platform-specific integration matters.

Sources

Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.

Related explainers

Run Whisper locally

Plan local Whisper transcription with the official VRAM table, whisper.cpp and Python routes, or WhisperJot's desktop models and honest download sizes.

faster-whisper

Read faster-whisper's CTranslate2 speed and memory benchmarks with their model and precision settings, Python requirements, and when this runtime fits.

WhisperKit

Learn how WhisperKit runs Whisper through Core ML and the Apple Neural Engine, which Apple platforms it targets, and what compressed model names mean.

See also: Whisper glossary definition · System requirements · Pricing