Higurashi releases Haru-4B, an open full-duplex voice model that listens and speaks at once
Haru-4B: A voice AI that can listen and speak at the same time
Higurashi released Haru-4B, a 4B full-duplex spoken dialogue foundation model built on Qwen3.5-4B that processes incoming and outgoing speech simultaneously, with weights on Hugging Face and code on GitHub. It uses 64-dimensional continuous acoustic latents at 7.5 Hz and a single-step MLP generation head, reaching about 65 ms end-to-end latency on an Apple M1 Max MacBook and about 15 ms per model step on an NVIDIA B300 GPU. The model was pretrained on over 40 million hours of real speech across more than 160 languages, including over 13 million hours of separated full-duplex dialogue, and is released as a pretrained base model for further fine-tuning and research.
Source:Hacker News · AI · blog.higurashi.ai