Underdog released Bark 3.0, and it's only a 163MB download. Runs locally w/Pytorch+CUDA or MLX On their Open ASR Leaderboard evaluation, Bark wins 🐶 Bark 3.0 → 4.94% WER → 163MB 🟢 NVIDIA Parakeet → 4.94% → 2.51GB 🌀 Whisper large-v3 → 6.04% → 9.26GB So Bark matches Parakeet's average accuracy while being 👉 ~15× smaller And beats Whisper while being 👉 ~57× smaller On recorded meetings check out these stats Bark → 7.82% WER Parakeet → 9.42% Whisper large-v3 → 13.63% And this isn't CUDA-server-only. 🏠 fully local 📱 phone-class 🍎 MLX + Core ML 🧠 Apple GPU / ANE 🪟 CPU + CUDA paths 🔓 CC-BY-4.0 I've checked and no GGUF. It uses its own compressed Bark format plus MLX/Core ML/Transformers runtimes. Bark 3.0 Plus is only 180MB and improves average WER to 4.76%. ⚠️ English only Link in ALT