higurashi-ai/Haru-4B
AI model · https://huggingface.co/higurashi-ai/Haru-4B
First seen by Oriel on .
AI summary
Haru-4B is a research foundation model for low-latency, multilingual, full-duplex spoken dialogue, generating audio and text continuously while listening.
- The model processes one 133-millisecond audio frame every step, producing its own audio and text at 7.5 Hz without turn-taking delays.
- A Qwen3.5-4B language backbone combines user audio, model audio, and model text, while a 4-layer MeanFlow MLP generates 64-dimensional audio latents in one step.
- Character cards control behavior and short reference clips control voices across English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean.
- The PyTorch implementation includes a web demo, MLX macOS app, and vLLM-Omni plugin, requiring approximately 12 GB of NVIDIA GPU memory.
- Haru-4B is not reliable for factual answers or assistant-style instruction following, and users must have rights to voices supplied as references.
#spoken dialogue #full-duplex ai #speech synthesis #multilingual ai
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 449, latest observation ; +308 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 3,372, latest observation ; +3,207 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | higurashi-ai/Haru-4B | https://huggingface.co/higurashi-ai/Haru-4B Read discussion |