List of OpenAI Whisper Checkpoints Variants
Subscribe
Official OpenAI Whisper has 12 unique checkpoint files , but the OpenAI Python package exposes 14 local model names because large and turbo are aliases for large-v3 and large-v3-turbo.<br>The real confusion comes from aliases, runtimes, and converted formats.<br>faster-whisper, CTranslate2, whisper.cpp, ggml, q8, q5, ONNX, Core ML, etc. are usually not new Whisper models . They are ways to run or compress the same models.<br>Official OpenAI Whisper checkpoints
#ModelTypeParametersNote1tiny.enEnglish-only39Mfastest English model2tinymultilingual39Mfastest multilingual3base.enEnglish-only74Mbetter than tiny4basemultilingual74Msmall multilingual baseline5small.enEnglish-only244Mgood lightweight English6smallmultilingual244Mgood lightweight multilingual7medium.enEnglish-only769Mstrong English model8mediummultilingual769Mstrong multilingual model9large-v1multilingual1550Moriginal large10large-v2multilingual1550Mimproved large11large-v3multilingual1550Mbest full Whisper model12large-v3-turbomultilingual~809Mfaster optimized large-v3
OpenAI’s README lists the six size families as tiny, base, small, medium, large, and turbo, with .en versions only for the first four families. It also states that turbo is optimized from large-v3 for faster transcription.<br>The OpenAI Python package accepts 14 names , because two names are aliases:<br>tiny.en<br>tiny<br>base.en<br>base<br>small.en<br>small<br>medium.en<br>medium<br>large-v1<br>large-v2<br>large-v3<br>large<br>large-v3-turbo<br>turbolarge = large-v3<br>turbo = large-v3-turbo
Use .en only when the audio is English. The .en models exist only for tiny, base, small, and medium. There is no large.en, large-v3.en, or turbo.en. OpenAI notes that .en models tend to do better for English, especially at tiny.en and base.en.<br>Use large-v3 when you want maximum local Whisper accuracy, but test by language/domain.<br>Use large-v3-turbo or turbo when you want much faster transcription with small accuracy loss.<br>Do not use turbo for translation.<br>OpenAI’s official openai-whisper package/CLI explicitly says turbo is not trained for translation and returns the original language even with --task translate.<br>Use medium, large, large-v2, or large-v3 for translation.
Name you seeWhat it isIs it a new Whisper model?large-v3official OpenAI Whisper checkpointYeslarge-v3-turboofficial OpenAI optimized checkpointYesturboalias for large-v3-turboNo, aliaslargealias for large-v3 in current OpenAI packageNo, aliaswhisper-1OpenAI hosted API model, powered by open-source Whisper V2API model ID, not a local checkpointgpt-4o-transcribenewer OpenAI transcription modelNot Whispergpt-4o-mini-transcribenewer cheaper transcription modelNot Whisperfaster-whisperPython implementation using CTranslate2NoCTranslate2 / ct2optimized inference format/runtimeNowhisper.cppC/C++ implementation using ggml filesNoggml-*whisper.cpp model file formatNoq5_0, q5_1, q8_0, int8quantized/compressed weightsNoONNXexported runtime formatNoCore MLApple runtime formatNoMLXApple Silicon ML runtime formatNodistil-whisperthird-party distilled Whisper-family modelsYes, but not OpenAI official
faster-whisper is not a separate model family. It is a faster implementation of Whisper using CTranslate2, and it can auto-download converted CTranslate2 versions of models. The project describes itself as a reimplementation of OpenAI Whisper using CTranslate2.<br>Current faster-whisper accepted names include:<br>tiny.en<br>tiny<br>base.en<br>base<br>small.en<br>small<br>medium.en<br>medium<br>large-v1<br>large-v2<br>large-v3<br>large<br>large-v3-turbo<br>turbo<br>distil-small.en<br>distil-medium.en<br>distil-large-v2<br>distil-large-v3<br>distil-large-v3.5The distil-* entries are Hugging Face Distil-Whisper models, not original OpenAI checkpoints.
whisper.cpp is a C/C++ runtime. Its downloadable ggml model list includes both normal and quantized files:<br>tiny<br>tiny.en<br>tiny-q5_1<br>tiny.en-q5_1<br>tiny-q8_0
base<br>base.en<br>base-q5_1<br>base.en-q5_1<br>base-q8_0
small<br>small.en<br>small.en-tdrz<br>small-q5_1<br>small.en-q5_1<br>small-q8_0
medium<br>medium.en<br>medium-q5_0<br>medium.en-q5_0<br>medium-q8_0
large-v1
large-v2<br>large-v2-q5_0<br>large-v2-q8_0
large-v3<br>large-v3-q5_0
large-v3-turbo<br>large-v3-turbo-q5_0<br>large-v3-turbo-q8_0That is 30 whisper.cpp downloadable file variants , but most are just quantized versions of the same underlying models. In the script, .en means English-only, q5_* / q8_0 mean quantized, and tdrz means TinyDiarize.<br>Simple understanding<br>Model = tiny/base/small/medium/large-v1/large-v2/large-v3/large-v3-turbo<br>Language suffix = .en only for tiny/base/small/medium<br>Runtime = openai-whisper / faster-whisper / whisper.cpp / transformers<br>Format = PyTorch / CTranslate2 / ggml / ONNX / Core ML / MLX<br>Compression = fp32 / fp16 / int8 / q5 / q8<br>Best accuracy: large-v3<br>Best speed/quality: large-v3-turbo or turbo<br>Best English small: small.en or medium.en<br>Best translation: large-v3, large-v2, large, medium<br>Avoid for translation: turbo<br>Best Mac local path: whisper.cpp + large-v3-turbo or large-v3<br>Best Python server: faster-whisper +...