FreeFlowDownload
Models · Speech model

Parakeet V3

By Yiftach FreemanUpdated 3 min read

Parakeet V3 is NVIDIA's 600-million-parameter speech model, run on the Mac's Neural Engine as a 484 MB CoreML model. It covers 25 European languages, detects which one you are speaking, and scores a 6.3% average English word error rate on the Open ASR Leaderboard — within a point of the most accurate model FreeFlow ships, at several times the speed. It is the default for a reason.

Parakeet V3 at a glance
Full nameNVIDIA Parakeet TDT 0.6B V3 (parakeet-tdt-0.6b-v3)
Download in FreeFlow484 MB
Languages25 European, with automatic detection
Average English WER6.3% (Open ASR Leaderboard, checked 2026-11-09)
Speed on Apple SiliconTypically under a second for a sentence, on the Neural Engine
Runs throughFluidAudio, an open-source CoreML runtime
Translate to EnglishNo
NeedsmacOS 14+, Apple Silicon

Why it is the default

Dictation lives or dies on the gap between letting go of the key and seeing the text. Parakeet is built as a token-and-duration transducer rather than an autoregressive decoder that emits one token at a time, so it finishes a sentence in a fraction of the time a Whisper model of similar accuracy takes, and it does it on the Neural Engine, which costs almost nothing in battery or fan. Its 6.3% average word error rate puts it within a point of Cohere Transcribe (5.4%) and ahead of every Whisper size FreeFlow ships (7.4% to 8%).

The 25 languages

Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian. It works out which one you are speaking from the audio. Not on the list: Hebrew, Arabic, Hindi, Japanese, Chinese, Korean — those need a Whisper model or Cohere Transcribe.

Where it is weaker

What FreeFlow does around it

The dictionary's replacements handle the names and identifiers no general model knows — Parakeet does not take vocabulary hints yet, so the heard-as → spelling form is the one to use with it. Modes add filler removal and, on macOS 26, Apple Intelligence formatting. And because macOS pages a 484 MB model out of memory during a long idle, FreeFlow re-reads the files in one pass while you are still speaking if the model has been idle ten minutes, so the first take after a break is as fast as the rest.

Questions people ask

Is Parakeet V3 better than Whisper for dictation?

For English and the 25 European languages, usually yes: lower word error rate than any Whisper size FreeFlow ships, and much faster. Whisper wins on languages outside those 25 and on translation.

Does Parakeet V3 do Hebrew or Arabic?

No. Hebrew needs Whisper; Arabic needs Whisper or Cohere Transcribe.

Does it send audio to NVIDIA?

No. The model runs as a CoreML file on your Mac; the only network use is the one-time download from Hugging Face.

Parakeet V2 or V3?

V3 unless you only ever dictate in English and want the last tenth of a point: V2 is English-only at 6.1%.

Sources

Try FreeFlow free for 14 days$99 once, three MacsNo card, no account. macOS 14+ · Apple Silicon.
Download now