Parakeet Redux
A ternary-quantized automatic speech recognition model developed by the Moondream team, based on NVIDIA's Parakeet 0.6B v3. It compresses model weights to three discrete values (−1, 0, +1), achieving approximately a sevenfold reduction in model size, and supports speech recognition across 25 European languages.
Timeline (last 90 days)
Parakeet Redux is a ternary-quantized speech-to-text model produced by the Moondream team based on NVIDIA Parakeet 0.6B v3, with the size compressed from approximately 1.2 GB to approximately 178 MB.
Moondream forces each weight in the Parakeet encoder into one of three values—−1, 0, or +1 (ternary quantization), requiring approximately 1.58 bits of storage per weight.
Ternary quantization is applied only to the encoder and not the decoder, which is one of the key reasons for the small accuracy loss.
Parakeet Redux's English word error rate (WER) increased from 6.26 to 6.55, its multilingual score improved from 11.62 to 10.56, and its long-audio score improved from 2.71 to 2.51.
Parakeet Redux supports 25 European languages and can automatically detect the language.
Parakeet Redux can run on CPU with a disk footprint of 178 MB, compared to the 1.6 GB of Whisper Large V3 Turbo (Whisper CPP version).
In CPU tests on an M2 MacBook Air, Redux achieved 38× real-time speed while Parakeet CPP achieved only 12×, a roughly 3× lead; however, on the Mac GPU the two are close (43× vs. 38/39×).
The 113× real-time speed claimed in the article title was measured on an AMD EPYC server with AVX512, not on a Mac.
On the noisy-audio benchmark, Parakeet Redux's error rate jumped from 6.72 to approximately 9, and the model card acknowledges that Redux is more prone to substituting words with phonetically similar ones.
Parakeet Redux's model weights are released under a Creative Commons license, but the driving engine Photon is not open source.
4 more timeline events
All Facts (14)
Parakeet Redux is a ternary-quantized speech-to-text model produced by the Moondream team based on NVIDIA Parakeet 0.6B v3, with the size compressed from approximately 1.2 GB to approximately 178 MB.
50%UnverifiedMoondream forces each weight in the Parakeet encoder into one of three values—−1, 0, or +1 (ternary quantization), requiring approximately 1.58 bits of storage per weight.
50%UnverifiedTernary quantization is applied only to the encoder and not the decoder, which is one of the key reasons for the small accuracy loss.
50%UnverifiedParakeet Redux's English word error rate (WER) increased from 6.26 to 6.55, its multilingual score improved from 11.62 to 10.56, and its long-audio score improved from 2.71 to 2.51.
50%UnverifiedParakeet Redux supports 25 European languages and can automatically detect the language.
50%UnverifiedParakeet Redux can run on CPU with a disk footprint of 178 MB, compared to the 1.6 GB of Whisper Large V3 Turbo (Whisper CPP version).
50%UnverifiedIn CPU tests on an M2 MacBook Air, Redux achieved 38× real-time speed while Parakeet CPP achieved only 12×, a roughly 3× lead; however, on the Mac GPU the two are close (43× vs. 38/39×).
50%UnverifiedThe 113× real-time speed claimed in the article title was measured on an AMD EPYC server with AVX512, not on a Mac.
50%UnverifiedOn the noisy-audio benchmark, Parakeet Redux's error rate jumped from 6.72 to approximately 9, and the model card acknowledges that Redux is more prone to substituting words with phonetically similar ones.
50%UnverifiedParakeet Redux's model weights are released under a Creative Commons license, but the driving engine Photon is not open source.
50%UnverifiedRunning pip install pulls in a package called Kestrel Kernels, whose license states that no usage rights are granted without signing a certain agreement.
50%UnverifiedThe core value of local speech transcription lies in privacy and offline operation, not in saving money.
50%UnverifiedDecision recommendations: Redux is worthwhile for pure-CPU or older machines and size-sensitive apps; users already leveraging M-series Mac GPUs achieve similar speeds with open-source runtimes and clearer licensing; noisy audio should use the original model.
50%UnverifiedOpen weights do not equal an open tech stack; Parakeet Redux's high performance is deeply dependent on the closed-source Photon runtime, and the community has no mature alternative implementation yet.
50%