whatmodel.sh / stt / Parakeet Ultra

Parakeet Ultra

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Parakeet Ultra is an open-weights speech-to-text model from Moondream, released September 2026. Its published size is 0.627B parameters. Licence: cc-by-4.0.

Catalogue identity notes

Moondream's post-trained moondream/parakeet-ultra (card: 'A post-trained version of parakeet-tdt-0.6b-v3. Same architecture, same tokenizer, same 0.6B parameters in full precision'; 25 European languages). Distinct from NVIDIA parakeet-tdt-0-6b-v3 (catalogued) and from its ternary sibling moondream/parakeet-redux. Repo created 2026-09-22; first and only commit 73175eb (2026-09-23), the upstream_commit the transcribe.cpp port records.

Model information sources

Release date ↗ · Parameter count ↗

Hugging Face safetensors metadata total 627,270,663 (current main 73175eb); transcribe.cpp catalog counts 627,265,543 for the converted port.

58.7±2.7#41 of 162 speech-to-text modelsfrom 5 benchmark results across 2 sources

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of Parakeet Ultra, out of 176 with a score, then more from Moondream.

#modelscorewhereRAMreleased
48GPT-Realtime-Whisper OpenAI59±3api $0.017/audio min2026-05
49Canary-1B NVIDIA59±3local~5.3GB2024-02
50Parakeet Ultra Moondream59±3local~0.7GB2026-09
51Parakeet TDT 1.1B NVIDIA59±2local~5.8GB2024-01
52Voxtral Mini 3B (2507) Mistral58±1api $0.0010/audio min +local~12GB2025-07
more from Moondream
64Parakeet Redux Moondream57±4local~0.2GB2026-09

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

other tests5 not in a task · 97% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-clean WER1.8% ✎▁···▁▁▂·▄▄██ beats 36/56avg 2.3%▇▇▇▇▇▇▇▇▇▇▇▇▇ 33%blunt
TED-LIUM long-form WER lab1.9% ✎█····█·█████ beats 6/6avg 4.6%▇▇▇▇▇▇▇ 19%sharp · small run
FLEURS 25 European languages, mean WER (NVIDIA set) lab9.6% ✎█····█·····█ beats 2/2avg 18%▇▇▇▇▇▇▇ 17%small run
Granite 5.0 report · Open ASR public 7-set mean WER lab5.8% ✎█···█······█ beats 2/2avg 11%▇▇▇▇▇▇ 15%small run
FLEURS English test WER4.3% ✎▁··▁▁▁▁▁▁▃█▁ beats 44/56avg 5.8%▇▇▇▇▇ 13%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~0.74 GB · from github.com/handy-computer
    git clone https://github.com/handy-computer/transcribe.cpp.git && cd transcribe.cpp && cmake -B build && cmake --build build && hf download handy-computer/parakeet-ultra-gguf parakeet-ultra-Q8_0.gguf --local-dir models/parakeet-ultra && build/bin/transcribe-cli -m models/parakeet-ultra/parakeet-ultra-Q8_0.gguf audio.wav

Official weights and documentation: huggingface.co/moondream ↗

$ how the score was worked out · Moondream parakeet-ultra and… 51%, transcribe.cpp model catalog… 47%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Moondream parakeet-ultra and parakeet-redux model cards (Benchmarks sections) · lab-reported3▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 51%CC BY 4.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports2▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 47%MIT
prior · pull toward the middle▇ 3%
weighted average= 58.7

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json