What is Moondream Parakeet Redux?
Parakeet Redux is a compressed version of NVIDIA's parakeet-tdt-0.6b-v3 speech-recognition model, released by Moondream (M87 Labs) on 22 September 2026. The architecture and tokenizer are unchanged, but every encoder weight is ternary, -1, 0 or +1, which cuts the weights from 1.2 GB to 178 MB. It transcribes the same 25 European languages, detects the language automatically, adds punctuation and capitalisation, and returns segment or word timestamps.
Moondream also trained a small voice-activity head into the weights, so long recordings are cut at pauses into pieces of at most 30 seconds without a separate VAD model. It runs in Photon, Moondream's inference engine, which computes directly from the packed weights on x86 (AVX-512 VNNI), ARM (NEON) and Apple GPUs (Metal); it is installed with pip install moondream. Streaming and live microphone input are supported.
The weights are under CC BY 4.0, like NVIDIA's original. A sibling, Parakeet Ultra, keeps full precision for GPUs and is further trained for accuracy.
Is Moondream Parakeet Redux worth it?
Parakeet Redux is the quickest way to run a strong open transcriber without a GPU, by Moondream's measurements. On eight cores of an AMD EPYC 9575F it transcribes LibriSpeech at 113 times real time, against 45 times for 8-bit parakeet.cpp and 42 for sherpa-onnx, and on a MacBook Air M2 it reaches 38 times on the CPU and 43 on the GPU. An hour of audio on that server takes about half a minute. All speed figures are Moondream's own, with every runtime on the same machine and the same audio.
Accuracy holds up better than the compression suggests. On the seven English sets of the Open ASR Leaderboard the average word error rate rises from 6.26% to 6.55%. On FLEURS across all 25 languages it falls from 11.62% to 10.56%, with large gains in Latvian (21.38 to 12.80) and Estonian, and losses in French (4.81 to 7.71), German and Russian. On 11 full TED talks it scores 2.51% against the original's 2.71%, which Moondream credits to its built-in pause detection.
The catch is noise. With MUSAN background noise mixed in, the average error rate climbs from 6.72% to 9.04%, and FLEURS German at 5 dB goes from 8.07% to 12.36%. Moondream says the ternary encoder substitutes similar-sounding words more often, though it does not drop or invent content more often. For call-centre audio or field recordings, Parakeet Ultra on a GPU, or NVIDIA's original, is the safer pick.
The other limit is the runtime. Photon is the only documented way to run the packed weights, so you cannot drop Redux into NeMo, ONNX or parakeet.cpp pipelines. The CC BY 4.0 licence is clean; the dependency on one vendor's engine is the trade-off for the speed.
Where it earns its place
A developer of a local meeting-notes app for Mac
Transcribe calls on the user's laptop without uploading audio
Photon runs on the Apple GPU at 43 times real time on an M2 MacBook Air, so an hour-long call takes under two minutes, and the 178 MB download is small enough to bundle. Word timestamps line up notes with the recording.
A platform engineer at a podcast-hosting company
Move transcription off GPUs that are needed for LLM work
At 113 times real time on eight EPYC cores, CPU servers can carry the transcription queue. Moondream cites a user whose Ryzen 9 9950X3D CPU beat 8-bit parakeet.cpp on an RTX PRO 6000 GPU for shorter clips; noisy uploads should still be spot-checked.
A public-sector translation team in the Baltics
Draft transcripts of Latvian, Lithuanian and Estonian recordings
On FLEURS, Redux cuts word error rate in Latvian from 21.38% to 12.80% and in Lithuanian from 21.09% to 17.27% against the original, so it is the better Parakeet for those languages even before speed.
Before you commit
- Setup
- pip install moondream (2.4.0 or later on the model card; the launch post uses 2.4.1), then open md.photon("moondream/parakeet-redux", device="cpu") and call transcribe on a WAV, MP3, FLAC or M4A file. The device can be cpu, mps or cuda.
- Integrations
- Python API with segment and word timestamps, streaming partial transcripts and async live input from a microphone or network stream; clip ranges and in-memory bytes are accepted. No NeMo, ONNX or ggml build of the ternary weights.
- Data and privacy
- Runs locally; audio stays on your machine. CC BY 4.0 allows commercial use with attribution to NVIDIA and Moondream.
- Cost at scale
- Free to run. On one 8-core EPYC 9575F chiplet at 113 times real time, 1,000 hours of audio is roughly nine hours of compute; no GPU is needed.
- Limits
- 25 European languages only; segments of at most 30 seconds, cut automatically. Live previews start after four seconds of audio and update every two seconds, with earlier text able to change.
What’s good
- 178 MB of weights against 1.2 GB for the original
- 113x real time on 8 x86 cores, 2.5x parakeet.cpp (Moondream)
- Better than the original on FLEURS and long-form talks
- Built-in pause detection; streaming and word timestamps
What’s not
- Error rate up from 6.72% to 9.04% in background noise
- Worse on French, German, Russian and Polish
- Runs only in Moondream's Photon engine
- All benchmarks are Moondream's own
Best for
Developers transcribing reasonably clean speech on CPUs, laptops or Macs who want Parakeet's 25 languages without a GPU.
Moondream Parakeet Redux alternatives
Moondream Parakeet Redux: common questions
- Is Moondream Parakeet Redux worth it?
- For transcription on a CPU or a Mac, Parakeet Redux trades a little English accuracy and a lot of noise robustness for a model one-seventh the size and more than twice as fast. We rate it 4.1 out of 5.
- How much does Moondream Parakeet Redux cost?
- Free weights under CC BY 4.0 (commercial use with attribution); runs locally through the free moondream Python package.
- What are the downsides of Moondream Parakeet Redux?
- Error rate up from 6.72% to 9.04% in background noise; worse on French, German, Russian and Polish; runs only in Moondream's Photon engine; all benchmarks are Moondream's own.
- Who is Moondream Parakeet Redux best for?
- Developers transcribing reasonably clean speech on CPUs, laptops or Macs who want Parakeet's 25 languages without a GPU.
- How much does Moondream Parakeet Redux cost for a team or in production?
- Free to run. On one 8-core EPYC 9575F chiplet at 113 times real time, 1,000 hours of audio is roughly nine hours of compute; no GPU is needed.
- Is Moondream Parakeet Redux safe for company data?
- Runs locally; audio stays on your machine. CC BY 4.0 allows commercial use with attribution to NVIDIA and Moondream.
- What are the best Moondream Parakeet Redux alternatives?
- The closest alternatives we have reviewed are Parakeet TDT 0.6B v3, Whisper large-v3-turbo and Voxtral 3 (Transcribe and TTS).
Answers checked Oct 31, 2026.
Make Moondream Parakeet Redux? Show this rating on your site
