Higurashi open-sources full-duplex speech model Haru-4B
Understand this story
Higurashi has released Haru-4B, a full-duplex continuous speech dialogue foundation model based on Qwen3.5-4B that can listen and speak simultaneously. The model is released as a pretrained foundation model for subsequent fine-tuning and research on real-time speech interaction. The model weights are available on Hugging Face, and the code is provided on GitHub. Higurashi said it plans to release a more detailed technical report and training code later.
AI overview based on source reports · Updated 2 days ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIHaru-4B: A voice AI that can listen and speak at the same time
Haru-4B is a full-duplex continuous spoken dialogue foundation model built on Qwen3.5-4B that processes incoming and outgoing speech simultaneously at 7.5 Hz, with about 65 ms end-to-end latency on a MacBook. It was pretrained on over 40 million hours of real speech across more than 160 languages, excluding LLM-generated and TTS-synthesized dialogues, and uses 64-dimensional continuous acoustic latents with single-step MeanFlow generation. Model weights are on Hugging Face and code on GitHub, released as a pretrained base model for further fine-tuning and research.
Activity for this story
- Current comparable activity
- 3
- Peak comparable activity
- 10Nov 19 06:00
- Change in the past 24 hours
- -49%
Trends compare the same fully observed participants, which may cover fewer participants than current activity. Move the pointer or tap the chart to inspect hourly activity; use the left and right arrow keys with a keyboard.