How widely is OpenAI Whisper used?

Whisper is widely used. In the 30 days to October 19, 2026, OpenAI's openai/whisper-large-v3-turbo was downloaded 6,094,388 times ¹. Counting offshoots, the Whisper family fills a large share of the top of Hugging Face's speech recognition downloads. It is not the single most downloaded model: that is a Japanese wav2vec2 model with 15,853,265 downloads ².

The data below comes from a ranking of the most downloaded speech recognition models on the Hugging Face Hub. The dataset holds 263 data points ³.

OpenAI's own Whisper models are downloaded in the millions

OpenAI publishes Whisper in several sizes, and the rows show demand across them. The family filter below lists the Whisper models published by OpenAI and others.

The turbo variant leads. Next, whisper-large-v3 has 3,755,126 downloads ⁴, and whisper-small has 3,510,611 downloads ⁵. Small models are popular because they are cheap to run: whisper-base has 1,676,272 downloads ⁶ and whisper-tiny has 818,386 downloads ⁷.

Re-packaged Whisper is where much of the use happens

Many developers do not run OpenAI's files directly. They use a version converted for speed or for a specific device, and those downloads are counted under other names.

• WhisperKit (argmaxinc/whisperkit-coreml), for Apple devices: 11,924,373 downloads ¹¹. This is more than any OpenAI Whisper release.
• faster-whisper (Systran/faster-whisper-small), built on the CTranslate2 library: 4,582,267 downloads ¹². The tiny version has 2,326,259 downloads ¹³.
• Distil-Whisper (distil-whisper/distil-large-v3), a smaller, faster distilled model: 589,897 downloads ¹⁴.
• Browser and GGUF builds, for example Xenova/whisper-tiny.en: 787,091 downloads ¹⁵, and handy-computer/whisper-large-v3-turbo-gguf: 479,865 downloads ¹⁶.

The family column is not filled for every Whisper-based row, so the Xenova and GGUF builds do not appear in the family table above. The full picture of Whisper use is larger than any one filter shows.

Other models are growing, but none matches Whisper's range

Newer engines now appear in the top 20. NVIDIA's Parakeet in an MLX build (mlx-community/parakeet-tdt-0.6b-v3) has 2,525,127 downloads ¹⁷, and Mistral's Voxtral Mini Realtime has 2,171,152 downloads ¹⁸. Both are well below the Whisper models. NVIDIA's own parakeet-tdt-0.6b-v3 has 626,738 downloads ¹⁹.

The wav2vec2 models high on the list are different in kind. Each is tuned for one language, such as jonatasgrosman's Japanese, Portuguese and Russian models. They serve narrow needs but are fetched by many pipelines.

What the numbers do and do not measure

• Downloads are not users. One download can serve thousands of people, and one automated job can download the same file many times. Pipelines that re-download a model on every run inflate counts.
• Whisper outside Hugging Face is missed. Apps that bundle Whisper, models served from GitHub releases, a vendor's CDN or Apple's Core ML packages are not counted. OpenAI's hosted transcription is not either.
• Tags are author-chosen. A few non-transcription audio models carry the speech recognition tag.
• Popularity is not accuracy. A high download count says nothing about how well a model works on your voice. For that, see our comparison of Whisper and Parakeet.

What this means for choosing a dictation tool

Whisper is a safe, well-supported default: tools, guides and fixes exist for almost every platform. But the data also shows the field moving toward models that run on your own device, in forms like WhisperKit, MLX and GGUF. If you dictate on a Mac or Windows PC, you do not have to pick a model yourself. Yap, which we make, runs dictation on your device, and you can compare it with other products on our market page.

To learn how these models work, read our overview of speech recognition and the guide to speech recognition models. To see every row with its maker, library and all-time downloads, open the full dataset.