---
title: "Moondream Shrinks Parakeet Redux Speech Recognition to 178 MB"
subtitle: "Moondream shrank NVIDIA's Parakeet speech model from 1.2GB to 178MB using ternary weights, hitting 113x realtime on CPU while matching accuracy."
url: "https://alphasignal.ai/news/moondream-shrinks-parakeet-redux-speech-recognition-to-178-mb"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-25T12:58:47.190Z"
updated: "2026-09-25T13:02:15.484Z"
type: "Model"
topics:
  - "Audio"
  - "Gpus"
subtopics:
  - "Speech To Text"
  - "Quantization"
  - "Edge Deployment"
upvotes: 1942
source_url: "https://huggingface.co/moondream/parakeet-redux"
comments: 0
truncated: true
---

# Moondream Shrinks Parakeet Redux Speech Recognition to 178 MB

> Moondream shrank NVIDIA's Parakeet speech model from 1.2GB to 178MB using ternary weights, hitting 113x realtime on CPU while matching accuracy.

# Parakeet Redux compresses speech recognition to 178 MB

Moondream has released [Parakeet Redux](https://huggingface.co/moondream/parakeet-redux), a 178 MB speech-to-text model derived from NVIDIA’s 1.2 GB `parakeet-tdt-0.6b-v3`. It retains the original architecture, tokenizer, and support for 25 languages while constraining every encoder weight to `-1`, `0`, or `+1`. The smaller representation enables fast local transcription on CPUs and Apple Silicon, with modest accuracy changes on clean speech and a larger regression in noisy conditions.

__Parakeet Redux at a glance__
| Model size      | 178 MB, down from 1.2 GB                |
| --------------- | --------------------------------------- |
| Encoder weights | Three values: \-1, 0, and +1            |
| Architecture    | Same as parakeet-tdt-0.6b-v3            |
| Languages       | English and 24 other European languages |
| License         | CC BY 4.0                               |

## Ternary weights cut memory traffic

Local inference on CPUs and Apple Silicon often depends on memory bandwidth because the processor must repeatedly fetch model weights. Packing each encoder weight into one of three possible states reduces that traffic substantially compared with 8-bit or 16-bit representations. The “1.58-bit” label comes from `log2(3)`, the information needed to represent three states before storage overhead.

Moondream’s [Photon](https://moondream.ai/photon) inference engine operates directly on the packed weights through hardware-specific kernels. It uses AVX-512 VNNI on supported x86 processors, NEON on ARM CPUs, and Metal on Apple GPUs.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [moondream/parakeet-redux - Hugging Face](https://huggingface.co/moondream/parakeet-redux)
- [Introducing Moondream Parakeet Redux and Parakeet Ultra](https://moondream.ai/blog/introducing-parakeet-redux-and-ultra)
- [NVIDIA Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3)