Haru 4B: open-weight Apache 2.0 full-duplex speech model that listens and talks simultaneously

thisguyknowsai · x · 2026-11-17

Higurashi released Haru 4B, an end-to-end full-duplex speech-to-speech model with Apache 2.0 weights that runs on 12GB of GPU memory. Unlike cascaded ASR-LLM-TTS systems, it processes audio directly, emitting one frame every 133ms so listening and speaking never block each other, enabling backchanneling, interruptions and natural turn-taking. Character is set via a text card and voice via a reference clip; it supports 9 languages. The model card cautions it's a research base model, not a reliable factual assistant, and can clone voices—only use voices you have rights to.

Original post →

More from Models

Models channel →