Original title: higurashi-ai/Haru-4B
A low-latency, multilingual full-duplex continuous speech dialogue foundation model released by Higurashi AI, for research purposes. Each frame is 133 milliseconds: the model reads one frame of user audio while simultaneously outputting one frame of its own audio and text; these three are summed and placed at the same position in the language model, so listening and speaking do not wait for each other. The audio consists of continuous vectors of 64 dimensions per frame, generated in one step by a small MLP head using MeanFlow, with no separate audio transformer and no discrete audio tokens. The language model backbone is the text component of Qwen3.5-4B, and the codec uses VibeVoice's causal acoustic tokenizer (24kHz). Personas are set via text character cards, and timbres are set using a short reference recording; it supports nine languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean.
Why it matters Listening and speaking occur synchronously at the same position in the language model, without waiting for the other party to finish; it can run on a single consumer-grade GPU, has a native Mac app, and the weights are Apache licensed
Upvotes 136 · Likes 149 · Downloads 165
Reasons
- Three implementations read the same weights: PyTorch (including a web demo), an MLX implementation on Apple silicon and a native macOS app, and a vLLM-Omni service plugin (multi-stream concurrency, low latency)
- Officially stated that the web demo can run on NVIDIA GPUs with approximately 12GB of VRAM; the Python interface sends and receives audio frame by frame, and there is also an offline command haru-inference
- The code repository and technical blog (in English and Japanese) are publicly available simultaneously
Usage
- git clone https://github.com/higurashi-ai/Haru,pip install -e ./haru, python -m haru.server --hf-repo higurashi-ai/Haru-4B, open localhost:8998 in a browser, select a character, provide a timbre, and start chatting directly
- No built-in timbres: you need to upload or record a clean, single-speaker recording between 3 seconds and 3 minutes; the official recommendation is to use wired headphones and a microphone, as Bluetooth adds more than 100 milliseconds of latency
- Character cards have a fixed format, generated using render_card in haru.prompt
Risks
- It is a pretrained foundation model: it can chat naturally, but the official documentation states it is not a reliable source of facts and does not follow instructions like an assistant
- It can mimic the voice of the reference recording; only use voices you have the right to use
- No public evaluation metrics are available, and the model card does not include comparisons with other full-duplex models; this entry has not been run