Parakeet Models Compared: V2, V3, Ultra, Redux

Which Parakeet Model to Use

What the Parakeet Model Names Mean

A name like parakeet-tdt-0.6b-v3 reads as decoder type, size, then version or language:

The NVIDIA Parakeet Model Family

NVIDIA trains Parakeet models with its NeMo toolkit. Each checkpoint pairs a FastConformer encoder, which converts audio into acoustic features, with a decoder that maps those features to text. NVIDIA has published more than a dozen Parakeet checkpoints on Hugging Face. These are the six newest general-purpose models:

Sources: NVIDIA model cards for V2, V3, and Unified, and Moondream's Ultra and Redux release.

Parakeet V2 vs V3

Parakeet TDT 0.6B V2 and V3 share the same design, with punctuation, capitalization, and word-level timestamps. V2 transcribes English only. V3 adds 24 more European languages and detects the spoken language automatically, with no language setting.

Word error rate (WER) counts missed, wrong, and extra words against a reference transcript, so lower is better. NVIDIA's model cards report these English averages on the Open ASR Leaderboard sets:

Parakeet Ultra and Redux: Moondream's V3 Variants

Moondream released both models on September 22, 2026. Both keep V3's architecture and tokenizer. A small voice activity detection (VAD) head lets Photon, Moondream's runtime, split long recordings at pauses into chunks of up to 30 seconds.

Moondream's English average excludes TED-LIUM and comes from a separate evaluation, so it is not directly comparable with NVIDIA's eight-set average. Business speech covers cleaned AMI meetings, VoxPopuli, and Earnings-22 calls; background noise mixes MUSAN noise into LibriSpeech and FLEURS at 10, 5, and 0 dB. None of the sets is dictation audio. To compare models for dictation, record a few quiet and noisy clips with names and numbers, run each model on them, and count the corrections.

Parakeet Ultra

Parakeet Ultra lowers word error rate in every test group. The largest relative drops are on full TED talks (28%) and 25-language FLEURS (18%), then background noise (13%); the TED result also reflects Photon's segmentation versus NeMo's long-audio path. On one NVIDIA B200 with 128 requests in flight, Photon reached an aggregate 9,743 times real time with Ultra, against 6,005 for V3 in NeMo at batch size 128. That comparison measures model and runtime together.

Parakeet Redux

Parakeet Redux restricts every encoder weight to -1, 0, or +1 (ternary, about 1.58 bits each), shrinking the weights from 1.2 GB at 16-bit to 178 MB. Background noise costs it the most accuracy, and although its FLEURS average is lower than V3's, it has a higher error rate in 11 of the 25 languages, including English, French, German, and Spanish. In Moondream's benchmark, Photon ran Redux at 113 times real time on eight AMD EPYC cores, 2.5 times as fast as parakeet.cpp on the same machine, and at 38 times real time on the CPU of an M2 MacBook Air.

Ultra and Redux on Apple Silicon

Fluid Inference's Core ML builds show the same pattern on an M-series Mac:

Redux takes about seven minutes to compile for the Neural Engine on first use on an M-series Mac; later loads take seconds. These results use 100 FLEURS clips per language, 15-second windows, and a simpler text normalizer, so they are not directly comparable with Moondream's.

Parakeet V3 Languages

Parakeet V3, Ultra, and Redux recognize the same 25 European languages:

Accuracy varies widely by language. In Moondream's FLEURS scoring, V3 ranges from 2.61% word error rate in Italian to 35.71% in Greek, and Greek stays the weakest on Ultra at 32.25%. Measured in percentage points, Ultra's largest gains come in Latvian, Slovenian, Lithuanian, and Maltese. For Chinese, Arabic, or other languages outside the list, use Whisper Large V3 Turbo, which covers about 99 languages, or Cohere Transcribe. The Parakeet vs Whisper comparison covers the two model families in detail.

Other Parakeet Models

NVIDIA also offers older English checkpoints, language-specific models, and a multitalker streaming model:

How to Run Parakeet Locally

The route depends on your hardware:

Community MLX, GGUF, and ONNX conversions of Ultra and Redux also appeared on Hugging Face within days of the release.

NVIDIA NeMo

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.ASRModel.from_pretrained(
    model_name="nvidia/parakeet-tdt-0.6b-v3"
)
output = asr_model.transcribe(["meeting.wav"])
print(output[0].text)

parakeet-mlx on a Mac

pip install parakeet-mlx -U
parakeet-mlx meeting.m4a --output-format txt

The command needs ffmpeg and loads mlx-community/parakeet-tdt-0.6b-v3 by default.

Ultra or Redux with Photon

Install the Moondream package with pip install --upgrade "moondream>=2.4.1", then transcribe a file with Redux on the CPU:

import moondream as md

with md.photon("moondream/parakeet-redux", device="cpu") as speech:
    result = speech.transcribe(audio="meeting.m4a")
    print(result["text"])

Set the model to moondream/parakeet-ultra and the device to "cuda" for Ultra on an NVIDIA GPU, or set the device to "mps" for Redux on an Apple GPU. Pass timestamps="word" for word timings.

Parakeet Models in Spokenly

Spokenly's default dictation models run in the cloud, and any of the local Parakeet models below can replace them. Local models are free, work offline once downloaded, and keep the audio on the device. On Mac and iPhone, Spokenly runs Fluid Inference's Core ML builds:

Ultra is larger than V3 because its Core ML encoder keeps 8-bit weights. Redux needs macOS 15 or iOS 18, and V2 and Unified need Apple silicon on a Mac; on Intel Macs, Spokenly recommends V3. On Windows, Spokenly runs Parakeet V3 (705 MB), V2 (696 MB), and Unified (697 MB) as GGUF builds, on the GPU through Vulkan when one is available. On Linux, V3 (456 MB) and V2 (451 MB) run on the CPU; the Linux speech-to-text guide covers the setup. Turn on Local Only Mode to block internet connections except model downloads while you dictate.

Parakeet FAQ

Read Next

• Whisper model sizes: Tiny to Large V3 Turbo
• Offline speech-to-text software
• How speech recognition works