Whisper · speech to text
Whisper large-v3-turbo (turbo)
OpenAI's README lists 809 M for this size. The exact count, 808,878,080, is Computed: I added up the tensor shapes in the official model.safetensors header on 19 Nov 2026. OpenAI's model-card.md says 798 M for turbo; the README and the file both say about 809 M.
OpenAI's GitHub README says the code and model weights are released under MIT. The Hugging Face repo metadata says apache-2.0 for every size except large-v3-turbo (mit). I follow the GitHub statement and flag the mismatch.
What it is
Whisper writes down speech. It is not a chat model, so this page has no context length and no KV memory math. OpenAI’s name for this file is large-v3-turbo. The README lists 809 M parameters. The exact count, 808,878,080, is Computed from the official safetensors header on 19 Nov 2026.
100 language codes: the 99 earlier ones plus Cantonese, added in large-v3 (Computed by counting LANGUAGES in whisper/tokenizer.py)
From the model card
Strengths
- large-v3 with the decoder cut from 32 layers to 4, then fine-tuned. OpenAI: 'faster transcription speed with a minimal degradation in accuracy.'
- The openai-whisper CLI uses turbo by default.
- Argmax's recommended model on iOS and macOS, in a compressed 626MB build.
Weaknesses
- Not trained for translation. With --task translate it returns the original language.
- On iPhone, first load can be slow while the Neural Engine compiles it (Community report, below).
- On Frankenstein's CPU (no AVX2) turbo was slower than medium: 74.38 s vs 60.64 s for a 161 s clip, because it keeps large-v3's full encoder. Measured on Frankenstein (HP ProLiant DL380p Gen8), 19 Nov 2026.
File size and memory
These are not the Frankenstein run. Each row keeps the label it came with.
| Figure | Value | Label |
|---|---|---|
| OpenAI 'Required VRAM' (openai-whisper, PyTorch) | ~6 GB | Vendor-reported Source |
| OpenAI relative speed vs large (A100, English) | ~8x | Vendor-reported Source |
| Hugging Face model.safetensors (float16) | 1.62 GB (1,617,824,864 bytes) | Vendor-reported Source |
| FP16 weights (params x 2 bytes) | 1.51 GiB | Computed params from the safetensors header |
| whisper.cpp ggml file | 1.5 GiB | Vendor-reported Source |
| whisper.cpp runtime memory | not published for turbo | Not yet measured Source |
| whisper.cpp ggml q5_0 file | 547 MiB | Vendor-reported Source |
Speed
| Model | Audio | File on disk | Wall time | Audio ÷ wall | Encoder | Peak RSS | JFK word errors |
|---|---|---|---|---|---|---|---|
| large-v3-turbo q5_0 | JFK sample, 11.0 s | 547 MiB | 11.40 s | 0.96× | 9.89 s | 801 MiB | 0.0% (0/22 words) |
| large-v3-turbo q5_0 | LibriVox Gettysburg Address, 161.2 s | 547 MiB | 90.87 s | 1.77× | 77.45 s | 908 MiB | — |
| large-v3-turbo | JFK sample, 11.0 s | 1,549 MiB | 10.11 s | 1.09× | 7.77 s | 1,802 MiB | 0.0% (0/22 words) |
| large-v3-turbo | LibriVox Gettysburg Address, 161.2 s | 1,549 MiB | 74.38 s | 2.17× | 61.34 s | 1,909 MiB | — |
Method, commit, and the exact command: methodology.
Measured on Frankenstein (HP ProLiant DL380p Gen8), 19 Nov 2026. A phone speed is not in this run. The Tesla P40 was not used.
Where it runs
Vendor-reported Compressed 626MB and 632MB builds (large-v3-v20240930) listed for A15 and newer; Argmax recommends the 626MB build "across iOS and macOS". This is Argmax’s list for WhisperKit, not a test I ran, and it names chips rather than phone models. See Run Whisper on your iPhone and Run Whisper on your computer.
Official downloads
These go to OpenAI or to the whisper.cpp GGML files. They are not store links.
- openai-whisper (pip), model name: turbo or large-v3-turbo
- Hugging Face openai/whisper-large-v3-turbo
- whisper.cpp ggml-large-v3-turbo.bin
Sources: openai/whisper README (model table) · openai/whisper model card · openai/whisper-large-v3-turbo model card · whisper.cpp models list (ggml file sizes) · WhisperKit device support (Argmax config.json) · large-v3-turbo model card