Haru GitHub repo: low-latency multilingual full-duplex spoken dialogue model

alex_verem · x · 2026-11-20

The inference repo for Haru-4B (higurashi-ai/Haru, Apache-2.0, 642 stars) is public. Haru is a low-latency, multilingual full-duplex spoken dialogue model: native full-duplex at 7.5 Hz with 133ms frames where user audio, model audio and text are summed into a single LM position; continuous audio generated in one MeanFlow sampling step via a small MLP head with no discrete audio tokens; text-card character control and reference-clip voice cloning. Ships three implementations (incl. MLX and vLLM) plus a web demo.

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

Original post →

More from Multimodal

Multimodal channel →