Haru-4B: open-source full-duplex voice model with 133ms frames, 9 languages

alex_verem · x · 2026-11-20

Haru-4B, the top trending model on Hugging Face, is an open-source (Apache-2.0) full-duplex spoken dialogue model: it listens and talks simultaneously at 7.5 Hz (133ms per frame), reacts to mid-sentence interruptions with 100ms latency, speaks 9 languages, clones a voice from a short sample, and runs locally on a Mac with MLX/vLLM implementations and a web demo.

Related event: Japanese Startup Open-Sources Full-Duplex Speech Model Haru-4B(3 posts)→

Original post →

More from Multimodal

Multimodal channel →