moondream/parakeet-redux

Moondream releases parakeet-redux, a 1.58-bit quantized variant of the Parakeet speech model that compresses weights to 178 MB. It runs at 113x real time on eight CPU cores using the Photon inference engine. Benchmarks show it beats the original on 25-language FLEURS data while remaining within 0.3 WER on English sets, enabling efficient local transcription without GPU acceleration.