Whisper model sizes: every model compared, and which fits your Mac

OpenAI's Whisper comes in six sizes: tiny (39M parameters), base (74M), small (244M), medium (769M), large (1550M) and turbo (809M). The four smallest also come as English-only models with a .en suffix. Bigger models are generally more accurate and slower. On a Mac, the practical limit is memory: whisper.cpp's own table puts medium at about 2.1 GB of memory and large at about 3.9 GB. For most people, small or medium is the starting point for English, and turbo or large-v3 when accuracy matters more than speed.

Every figure below comes from the project that published it, opened on 1 October 2026. We have not run a controlled benchmark of any size. Benchmark numbers describe the test audio those projects used, not your voice, accent, microphone or vocabulary.

What are all the Whisper model names?

The openai/whisper README lists the sizes, names, memory and relative speed. The Hugging Face model card for large-v3-turbo adds the separate large versions.

How to read this table:

• Relative speed is against the large model, on an NVIDIA A100 GPU. The README says the speeds were "measured by transcribing English speech on a A100, and the real-world speed may vary significantly depending on many factors including the language, the speaking speed, and the available hardware." It is not a Mac figure.
• Required VRAM is for OpenAI's own PyTorch code on a graphics card. Mac-focused projects publish lower figures (next section).
• "large" means large-v3 today. In OpenAI's package code, the name large downloads the same file as large-v3, and turbo downloads large-v3-turbo.
• English-only models. The README says the .en models "tend to perform better, especially for the tiny.en and base.en models," and that the difference "becomes less significant for the small.en and medium.en models."
• Languages. The multilingual models cover many languages, and the large-v3 card lists 99 language codes. The README warns that "Whisper's performance varies widely depending on the language."

The whisper-large-v3-turbo card explains turbo: "it's the exact same model, except that the number of decoding layers have reduced from 32 to 4." Large-v3 vs turbo covers that trade-off and the translation caveat.

How much disk and memory does each size need on a Mac?

The whisper.cpp README, the C/C++ version that calls Apple Silicon "a first-class citizen", publishes its own figures for its model files:

These are whisper.cpp's numbers for its own file format. Other tools store the same model differently, so file sizes differ. On Hugging Face, for example, the large-v3 model.safetensors file is listed at about 3.09 GB and the large-v3-turbo file at about 1.62 GB on each repository's Files tab. Always check the size a tool shows before you download.

Which Whisper size should you pick for your Mac?

There is no published Mac-by-Mac chart from OpenAI, so use the memory figures above and leave room for the apps you keep open.

1. 8 GB of memory. Start with small or small.en. Medium fits on whisper.cpp's figures, but you are sharing memory with your browser and editor. Tiny and base are useful for testing a setup, and WhisperKit's README describes tiny as "Smallest size, lowest accuracy. Only recommended for development & debugging."
2. 16 GB or more. Medium, turbo and large-v3 become realistic. Try turbo first if you transcribe long files and want speed; try large-v3 if your language is one where turbo is weaker.
3. English only. Use the .en version of tiny, base, small or medium. There is no English-only large or turbo.
4. Another language, or several. Use a multilingual model. If you need speech translated into English, do not use turbo (see below).
5. Live dictation rather than files. Latency matters more than for a file you leave running. A smaller model that returns text quickly can beat a larger one that makes you wait. Test both.

Whatever you choose, test it on your own speech. Size predicts cost better than it predicts accuracy on your voice.

What do quantised and Core ML versions change?

You rarely run OpenAI's original files on a Mac. You run a converted copy.

• Quantised models (whisper.cpp). The README says whisper.cpp "supports integer quantization of the Whisper ggml models" and that "Quantized models require less memory and disk space and depending on the hardware can be processed more efficiently." Its example converts base.en with the Q5_0 method. Quantisation stores the weights with fewer bits, so it can change output slightly. Compare transcripts before you rely on a quantised file.
• Core ML (whisper.cpp). On Apple Silicon, the README says the encoder "can be executed on the Apple Neural Engine (ANE) via Core ML," which "can result in significant speed-up - more than x3 faster compared with CPU-only execution." It also warns that "The first run on a device is slow, since the ANE service compiles the Core ML model."
• WhisperKit. Argmax's argmax-oss-swift README lists Core ML versions of the Whisper sizes, including compressed ones. It recommends large-v3-v20240930_626MB, a compressed Large v3 Turbo, "across iOS and macOS for maximum accuracy," and the uncompressed large-v3-v20240930_turbo "on macOS for maximum speed and accuracy." It also says WhisperKit "automatically downloads the recommended model for the device if not specified."

The model's parameter count stays the same. What changes is the file size, the memory used, which chip does the work, and possibly small differences in output.

Where does SuprFlow fit?

SuprFlow, the Mac dictation app that publishes this guide, uses a WhisperKit (Whisper) model in version 1.0.0. It offers Small and Medium, each in multilingual and English-only variants, and downloads the one you choose during setup. It does not offer tiny, base, large or turbo. Compare SuprFlow speech models shows how to test Small against Medium on your own voice.

Questions people ask

How big is Whisper large-v3?

It has 1550M parameters, per OpenAI's README and model card. The README lists about 10 GB of graphics memory for OpenAI's own code. whisper.cpp lists its large model file at 2.9 GiB on disk and about 3.9 GB in memory. The Hugging Face model.safetensors file is about 3.09 GB.

Is there an English-only Whisper large or turbo model?

No. OpenAI's README and model card list English-only versions only for tiny, base, small and medium.

Is a bigger Whisper model always more accurate?

Usually on benchmarks, not always for you. The large-v3 card says the models "exhibit disparate performance on different accents and dialects," and OpenAI's README says performance "varies widely depending on the language." Run the same recording through two sizes and count the errors.

Which Whisper model is fastest on a Mac?

Tiny is the smallest and fastest in OpenAI's table, and turbo is listed at ~8x the large model's speed. Those figures come from an A100 GPU, so measure on your own Mac.

Related guides

• How to run Whisper on a Mac
• Whisper large-v3 vs turbo
• Whisper vs faster-whisper vs WhisperX
• Parakeet vs Whisper
• Speech to text accuracy and word error rate
• Compare SuprFlow speech models

Sources

• OpenAI — whisper GitHub repository and README
• OpenAI — whisper package model list (whisper/__init__.py)
• OpenAI — whisper-large-v3 model card on Hugging Face
• OpenAI — whisper-large-v3-turbo model card on Hugging Face
• OpenAI — whisper-large-v3 files on Hugging Face
• whisper-large-v3-turbo files
• ggml-org — whisper.cpp GitHub repository and README
• Argmax — argmax-oss-swift (WhisperKit) GitHub repository and README