163MB open-source speech model Bark 3.0 beats Whisper (57x bigger) and NVIDIA Parakeet (15x bigger)
Scobleizer · x · 2026-10-21
Underdog released Bark 3.0, a tiny 163 MB open-source (CC-BY 4.0) speech recognition model that runs fully on-device on laptops and phones.
- Benchmarks: 4.94% average word error rate across the Open ASR Leaderboard's seven English test sets, beating NVIDIA Parakeet (9.42%, 15x larger) and OpenAI Whisper large-v3 (13.63%, 57x larger). It transcribes 1 minute of audio in under 0.25s on a MacBook, up to 595x faster than real time.
- Family: Bark 3.0 Plus (180MB, 4.76%), Pro (1.25GB, 4.29%, Hugging Face only), and Clean, which deliberately strips fillers, false starts and repeats for dictation and meeting notes.
- Shipping: integrated into the Underdog app for private on-device meeting notes and dictation.
More from Models
- Reflection's open-weight Beam 1.0 (501B-A23B) free on Nous Portal for a week — alexcovo_eth · 2026-10-21
- Karpathy: you know RL is working when models stop thinking in English — joeddav · 2026-10-21
- Devin Fusion's dual-model setup scores 68 on Coding Agent Index at 31% lower cost than Claude Code — ArtificialAnlys · 2026-10-21
- Perplexity open-sources pplx-decider models at 1/10th the cost of GPT 6 Luna — Balance- · 2026-10-21
- Ascend 950DT hits 1/5-1/7 of GB200 decode throughput on DeepSeek-V4.1, but EP optimization may close the gap — teortaxesTex · 2026-10-21
- Gemini 4 Carbon enters early access, spotted in LMArena, stronger at 3D than Argon — lyraxana · 2026-10-21