# Whisper large-v3 turbo Canonical page: https://flit.fyi/glossary/whisper-large-v3-turbo [Glossary](https://flit.fyi/glossary) / [How it hears you](https://flit.fyi/glossary#recognition) Also: Whisper turbo, large-v3-turbo Published October 8, 2026 Definition Whisper large-v3 turbo is a speech recognition model from OpenAI, released on 1 October 2024. It is Whisper large-v3 with the decoder cut from 32 layers to 4, which makes it much faster at a small cost in accuracy. Whisper comes in sizes, from tiny to large, and for years the choice was between a small model that was quick and a large one that was accurate. Turbo changed that. OpenAI kept the large encoder, the part that listens, and replaced the large decoder, the part that writes, with one as shallow as the tiny model has. The result has 809 million parameters against 1,550 million for large-v3. OpenAI describes it as offering faster transcription speed with a minimal degradation in accuracy. ## What changed from large-v3 OpenAI’s announcement puts it in one sentence: turbo is an optimized version of large-v3 with only 4 decoder layers, the same as the tiny model, down from the 32 in the large series. The model card calls it a fine-tuned version of a pruned large-v3. After the decoder was cut down, the model was trained for two more epochs on the same multilingual transcription data as large-v3. Translation data was left out, which matters for the next section. ## How fast and how accurate it is OpenAI’s own table lists turbo at about 8 times the speed of the large model, and about 6 GB of video memory against about 10 GB. Those figures were measured transcribing English speech on an A100 graphics card, so they describe the ratio, not what a particular Mac will do. On accuracy, OpenAI says that across languages turbo performs similarly to large-v2, with larger degradation on some languages, and it names Thai and Cantonese. ## What it cannot do Turbo shares the limits of every Whisper model, and adds one of its own. - It does not translate. OpenAI states that turbo is not trained for translation tasks and returns the original language even when asked to translate into English. - It can hallucinate. The model card says predictions may include text that was not actually spoken. - It can repeat itself. The card says the architecture makes it prone to generating repetitive text. - It is uneven across languages, in the card’s own words. - It is not built for live use. The card says Whisper models cannot be used for real-time transcription out of the box. ## Size, languages and license The published weights are about 1.6 GB in 16-bit form. Smaller builds exist because the weights can be quantized, which trades a little accuracy for a smaller download. The model card lists 99 languages. The license is MIT for both the code and the weights, so it can be used and shipped freely. ## Running it on a Mac There is no single way. OpenAI publishes the model for Python. Other projects convert the same weights to run on Apple silicon, either on the graphics chip or through Apple’s own frameworks, and those are what Mac apps build on. Because Whisper reads a chunk of audio before it writes, it suits finished recordings and dictation that appears when you stop. Tools that show words while you are still speaking chunk the audio themselves, or use a streaming model for that part. In Flit Whisper large-v3 turbo is one of the voice models you can choose in Flit. It is a download, and it runs on your Mac. [The languages Flit supports](https://flit.fyi/#languages) ## Questions people ask. ### Is Whisper large-v3 turbo better than large-v3? It is faster, not more accurate. OpenAI describes it as offering faster transcription with a minimal degradation in accuracy, and lists it at about 8 times the speed of the large model. ### How big is Whisper large-v3 turbo? It has 809 million parameters. The published 16-bit weights are about 1.6 GB, and quantized builds are smaller. ### Can Whisper turbo translate? No. OpenAI states that the turbo model is not trained for translation tasks and returns the original language even if translation is requested. ### How many languages does Whisper large-v3 turbo support? Its model card lists 99 languages. OpenAI notes that accuracy is uneven across them. ### Is Whisper large-v3 turbo free to use? Yes. OpenAI releases Whisper’s code and model weights under the MIT License. ### Does Whisper turbo work offline? Yes. The weights are a file. Once it is downloaded, the model runs on your own machine with no connection. ## Related terms - [Whisper](https://flit.fyi/glossary/whisper) - [Parakeet](https://flit.fyi/glossary/parakeet) - [Quantization](https://flit.fyi/glossary/quantization) - [Hallucination](https://flit.fyi/glossary/hallucination) - [Automatic speech recognition](https://flit.fyi/glossary/automatic-speech-recognition) ## Your voice was always faster. 7 days free, then $19 USD once. No account, no subscription, nothing uploaded. [Download for Mac](https://flit.fyi/download) [Buy now · $19 USD](https://flit.fyi/pricing#personal) macOS 14 or later · Apple silicon