Whisper is the open-source speech-recognition model family offered in sizes from tiny through large, with turbo as a faster transcription option. Model size changes parameter count, required VRAM and relative speed. The official table below is a starting point for choosing a checkpoint; its GPU measurements are not universal requirements for every local runtime.
| Size | Parameters | English-only model | Multilingual model | Required VRAM | Relative speed |
|---|---|---|---|---|---|
| tiny | 39 M | tiny.en | tiny | ~1 GB | ~10x |
| base | 74 M | base.en | base | ~1 GB | ~7x |
| small | 244 M | small.en | small | ~2 GB | ~4x |
| medium | 769 M | medium.en | medium | ~5 GB | ~2x |
| large | 1550 M | N/A | large | ~10 GB | 1x |
| turbo | 809 M | N/A | turbo | ~6 GB | ~8x |
Values reproduced from the official Whisper README. Relative speeds measure English speech on an A100; language, speaking speed and hardware affect actual results. The large row does not identify the checkpoint behind the alias.
The official README reports relative speed for English speech on an A100. It also says results vary with language, speaking speed and available hardware. Use its VRAM column to compare the listed models within that setup, then check the runtime you plan to use. A converted GGML file's disk size answers a different question from GPU memory needed during recognition.
The .en suffix identifies an English-only model. The README says these variants tend to perform better for English-only applications, especially tiny.en and base.en. Large and turbo have no .en variant in the official table. If you need multilingual transcription, select a multilingual checkpoint instead of assuming that an English-specific model will detect and transcribe another language.
Start by deciding whether you need transcription or translation, which languages you use, and which runtime will load the weights. Turbo is optimized for transcription and is not trained for translation. The large-v3 card explicitly lists 1550 M parameters; large-v3-turbo has 809 M. Test candidate models with representative speech rather than treating the relative-speed column as an accuracy ranking.
WhisperJot uses Whisper in Jot Local Pro, its local engine for long-form accuracy. On macOS it runs through WhisperKit with large-v3-turbo as the default model. On Windows and Linux it loads whisper.cpp-format GGML models through a .NET runtime. Auto mode picks between Jot Local Pro and the Parakeet-based Jot Local per utterance. Both local engines work offline after their one-time model download.
Explore WhisperJot →The official table lists tiny and base at ~1 GB required VRAM. Those are the lowest entries in that table. This is a reference for the documented implementation; a different runtime, conversion or quantization can have a different memory profile.
No. The official table marks the English-only entries for large and turbo as N/A. It lists English-only variants for tiny, base, small and medium. Choose the multilingual version when your audio is not English. The suffix is a language-scope choice, not a runtime name.
No. Large-v3 has 1550 M parameters, while large-v3-turbo has 809 M. The official table lists ~10 GB VRAM for large and ~6 GB for turbo. Disk downloads depend on the format and compression used by the runtime. The smaller decoder is part of turbo's design.
Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.
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.
Understand Whisper large-v3-turbo: fewer decoder layers, reference speed, translation limits, compressed variants and its role in WhisperJot Jot Local Pro.
Plan local Whisper transcription with the official VRAM table, whisper.cpp and Python routes, or WhisperJot's desktop models and honest download sizes.
See also: Whisper glossary definition · System requirements · Pricing
