The speech models
FreeFlow ships six on-device speech models. Parakeet V3 (484 MB, 25 European languages, 6.3% average word error rate) is the default and the fastest. Cohere Transcribe is the most accurate at 5.4% but needs 2.2 GB and macOS 15. The three Whisper models cover 99 languages, including Hebrew, Arabic and CJK, and can translate to English as you speak.
| Model | Download | Languages | Avg WER | Speed | Notes |
|---|---|---|---|---|---|
| Parakeet V3 (default) | 484 MB | 25 European | 6.3% | Fastest | NVIDIA. Detects the language itself; the mode's language keeps it in the right alphabet |
| Parakeet V2 | 465 MB | English only | 6.1% | Fastest | A fraction sharper on English, nothing else |
| Cohere Transcribe | 2.2 GB | 14 | 5.4% | Slow (1.5–3 s a take) | Most accurate. No auto-detect. macOS 15+ |
| Whisper Small | 970 MB | 99 | ~8% | Fast | Lightest route to Hebrew, Arabic, CJK; noticeably less accurate |
| Whisper Large Turbo | 1.6 GB | 99 | 7.8% | Medium | Best choice for Hebrew; decoder trimmed for speed |
| Whisper Large | 3.1 GB | 99 | 7.4% | Slowest | The full model |
Which one to pick
- Most people: Parakeet V3. Near-instant on Apple Silicon, within a point of the most accurate model, and it works out which of its languages you are speaking.
- English only, every tenth of a point counts: Parakeet V2. Same speed, 6.1%.
- Accuracy over speed, in one of its 14 languages: Cohere Transcribe. English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek, Arabic, Japanese, Chinese, Korean, Vietnamese. It transcribes the language the mode names — there is no auto-detect — and a take costs a couple of seconds.
- Hebrew, or anything outside the lists above: a Whisper model. Large Turbo is the sweet spot. Whisper also offers translate-to-English, which types English while you speak something else.
Where they run
All six are CoreML models, so they run on the Mac's own chip. Settings → General → Speech model compute chooses between the Neural Engine, CPU & GPU, or Automatic; Automatic is right unless you have a reason. A change applies from the next dictation. If a model ever fails to prepare on the Neural Engine at launch, FreeFlow notices on the next launch and puts the setting back to Automatic rather than refusing to start.
The first take after a long idle used to be slow: macOS quietly pages a model's weights out to make room for other apps, and pulling 2 GB back one page at a time took up to fifteen seconds. Since 1.5.2 FreeFlow re-reads the model's files in one pass while you are still speaking if the model has not been used in ten minutes, so the weights are back before you let go of ⌥ — about a third of a second for Cohere's 2 GB.
Per mode, not per app
Each mode keeps its own model and language, so a “Hebrew notes” mode on Whisper Large Turbo and a “Code” mode on Parakeet V2 can sit side by side in the menu bar. Switching modes switches models; a model that is not yet loaded loads on first use.
Bring your own
Models → Add accepts any Hugging Face repository in WhisperKit's CoreML layout — a fine-tune, a distilled variant, a model that did not exist when FreeFlow shipped — or a model folder already on disk. It appears in the list beside the built-ins and can be assigned to any mode. Downloads can be cancelled with the × beside the progress bar, and a failed download says why.
Questions people ask
Which model is the most accurate?
Cohere Transcribe, at 5.4% average word error rate, in its 14 languages. Parakeet V3 is within a point of it at 6.3% and is many times faster, which is why it is the default.
Does Parakeet V3 do Hebrew?
No. Its 25 languages are European and do not include Hebrew, Arabic or the East Asian languages. Hebrew needs a Whisper model; Arabic, Japanese, Chinese and Korean need Whisper or Cohere Transcribe.
How much disk do I need?
484 MB for the default. Up to about 8 GB if you install every model, which nobody needs to.
Can I run two models at once?
One model is resident at a time per engine; switching modes swaps them, and a cross-engine switch unloads the previous one so two multi-gigabyte pipelines never sit in memory together.