Ternary-compressed Parakeet speech model shrinks 1.2GB to 178MB, runs 113x realtime on CPU
TheZachMueller · x · 2026-09-22
vikhyatk released Parakeet Redux, a ternary speech-to-text model created by compressing NVIDIA's Parakeet from 1.2GB down to 178MB.
- Runs at 113x realtime on CPU
- Beats the base model on the 25-language FLEURS benchmark
- English WER stays within 0.3 of the original
More from Models
- Frontier coding capability has plateaued since Opus 4.8 as OSS models close in at 10-50x lower cost — Yuchenj_UW · 2026-09-22
- Grok 4.7 ranks #2 on EEBench, beating Claude Fable 5.1 and Opus 5 on real-world EE tasks — XFreeze · 2026-09-22
- Grok 4.7 lands in Agent Arena; community votes on millions of real agentic tasks — arena · 2026-09-22
- Scoble: Grok 4.7 turns the frontier model race into an economics test — Scobleizer · 2026-09-22
- Grok 4.7 is out; Agent Arena polls where it will land on its score trend — therealdanvega · 2026-09-22
- CoT monitors catch reward hacking, but optimizing against them teaches models to hide it — gordic_aleksa · 2026-09-22