Technical explainer · Verified 2026-09-05

Whisper large-v3-turbo
speed and tradeoffs

Whisper large-v3-turbo is a pruned and fine-tuned version of Whisper large-v3 for faster transcription. It reduces the decoder from 32 layers to 4 and has 809 M parameters. The smaller decoder trades some recognition quality for speed, and translation data was excluded from its fine-tuning, so it is a transcription model rather than a translation model.

A smaller decoder, with a specific speed reference

The turbo card describes pruning the decoder from 32 layers to 4 while retaining the large-v3 starting point. The official README lists ~8x relative speed and ~6 GB required VRAM for turbo. Those speed measurements use English speech on an A100. They are useful for comparing the README rows, but they do not predict elapsed time on a particular laptop.

Translation and language tradeoffs

The release announcement says turbo performs similarly to large-v2 across languages, with larger degradation on languages such as Thai and Cantonese. It also explains that translation data was excluded during fine-tuning. The README directs translation users to the multilingual non-turbo models. Treat turbo as a transcription choice and test the languages you actually need, rather than assuming large-v3-equivalent quality throughout.

Download size depends on the representation

The Transformers-format model.safetensors file is 1.62 GB. The whisper.cpp table lists large-v3-turbo at 1.5 GiB and its q5_0 version at 547 MiB. WhisperKit also lists compressed variants with 626MB and 547MB in their names. These labels describe different representations; a size in a folder name is not a measurement of runtime memory or a universal app download requirement.

How WhisperJot uses this

WhisperJot's macOS Jot Local Pro is WhisperKit-based. Its default large-v3-turbo download is approximately 1.5 GB; selectable alternatives are compressed large-v3-turbo at approximately 626 MB and small.en at approximately 217 MB. Non-English transcription requires large-v3-turbo because the .en option is English-only. WhisperKit's full catalog offers additional models beyond this product selection. Jot Local Pro runs on-device by default and works offline after the model download.

Explore WhisperJot →

Frequently asked questions

Can Whisper turbo translate speech into English?

Turbo is not trained for translation tasks. The README directs users who need translation to multilingual tiny, base, small, medium or large models. Being multilingual for transcription does not mean turbo has the same translation training as those models. Choose accordingly.

Is large-v3-turbo as accurate as large-v3?

The turbo model card describes a minor quality degradation. The announcement says performance is similar to large-v2 across languages, with larger degradation on Thai and Cantonese. That makes language-specific testing more useful than assuming identical accuracy because both names contain large-v3.

Why are there different turbo download sizes?

The same checkpoint can be converted and compressed for different runtimes. Transformers weights, GGML files and Core ML variants use different representations. WhisperKit folder labels also distinguish variants, so compare the actual runtime and model selection before comparing their size labels.

Sources

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

Related explainers

Whisper large-v3

Learn what Whisper large-v3 changed from large-v2, how its language coverage and error-reduction claim are stated, and which license the project cites.

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.

whisper.cpp

Explore whisper.cpp, its GGML model disk sizes, integer quantization and Apple acceleration options, plus how WhisperJot uses GGML on Windows and Linux.

See also: Whisper glossary definition · System requirements · Pricing