Parakeet Redux fits 1.2 GB of speech recognition into 178 MB

Moondream is extending NVIDIA's Parakeet speech-to-text model with two variants: Redux compresses its weights to 178 MB for faster local inference on CPUs and Macs, while Ultra keeps full-precision weights and targets lower transcription error rates on GPUs.

Three values to compress the weights

Running a speech recognition model locally is not only a matter of compute. On CPUs and Apple Silicon, moving weights between memory and the processor can become the primary inference bottleneck.

Moondream Parakeet Redux reduces that traffic by applying ternary quantization to Parakeet's encoder: each weight can take only three values, -1, 0, or +1. The model shrinks from 1.2 GB to 178 MB while retaining the same architecture, tokenizer, and support for 25 languages.

Moondream's Photon runtime works directly with that packed representation using kernels optimized for different hardware architectures. On a MacBook Air M2, the team reports processing 38 seconds of audio per second on CPU and 43 on GPU. Using eight cores of an AMD EPYC 9575F, throughput reaches 113 seconds of audio per second. These performance figures come from Moondream's own benchmarks. Heavy compression with a limited accuracy tradeoff

Compressing the weights this aggressively raises an immediate question about accuracy. Moondream measures it using Word Error Rate, or WER, where lower scores indicate fewer transcription errors.

Across seven English test sets, Redux reaches an average WER of 6.55%, compared with 6.26% for the original Parakeet. On the 25-language FLEURS evaluation, however, the compressed version improves from 11.62% to 10.56%. On long-form TED-LIUM recordings, it also moves from 2.71% to 2.51%.

The tradeoff becomes more apparent with background noise, where WER rises from 6.72% on the original model to 9.04% with Redux. For these recordings, Moondream points users toward its second variant, Parakeet Ultra. Ultra keeps the size and targets accuracy

Parakeet Ultra takes the opposite direction. Its weights remain at full precision, while additional post-training focuses on reducing transcription errors, with GPU execution in mind.

According to Moondream's evaluations, Ultra reaches 5.80% WER across the seven English test sets, 9.55% on FLEURS, 5.79% on business speech, and 5.82% with background noise. Across eleven complete TED-LIUM talks lasting 10 to 20 minutes, it reaches 1.94%, compared with 2.71% for the original Parakeet.

These results are published by Moondream and should not be treated as an independent comparison against the broader speech-to-text market. Audio segmentation moves inside the model

Both variants also integrate voice activity detection. For longer recordings, Photon identifies pauses and automatically splits audio into segments of up to 30 seconds, removing the need to run a separate VAD model.

Transcription can run on files or live audio streams, with timestamps at either segment or word level. In streaming mode, Photon begins returning previews after four seconds of audio and updates them every two seconds as additional context becomes available.

Redux weights are available through Hugging Face under the CC-BY-4.0 license, matching NVIDIA's original Parakeet model. Redux primarily targets CPUs and Apple Silicon, while Ultra retains full precision for GPU workloads.