✦
Summary
Haru 4B is higurashi's end-to-end full-duplex speech-to-speech model, released in November 2026 as an open-weight pretrained base model for research. It listens and speaks at the same time, handling backchannels, interruptions and overlapping speech the way a real conversation works, instead of the turn-taking of a cascaded pipeline. It is built on Qwen3.5-4B with a multi-stream architecture synced at 7.5 Hz, and its open weights and code are Apache 2.0 licensed.
Key Features
Full-duplex listening and speaking at the same time
Backchanneling, interruption and overlapping-speech handling
Multi-stream architecture synced at 7.5 Hz with about 133 ms frames
Multilingual: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean
Character control via a text card plus voice cloning from a 3-second to 3-minute reference clip
Continuous speech tokenization with the VibeVoice tokenizer at 24 kHz
Three implementations: PyTorch, MLX with a native macOS app, and a vLLM-Omni serving plugin
Web demo server and Python API with DuplexSession and an offline inference CLI
About Haru 4B
The model supports nine languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. Character control works through a text card plus voice cloning from a reference clip of 3 seconds to 3 minutes, and speech is tokenized continuously with the VibeVoice tokenizer at 24 kHz. Three implementations ship: PyTorch, MLX with a native macOS app, and a vLLM-Omni serving plugin, plus a local web demo server and a Python API with DuplexSession and an offline inference CLI.
Haru 4B has no hosted product: users clone the GitHub repo, download the weights from Hugging Face as higurashi-ai/Haru-4B, and run the demo locally with python -m haru.server on localhost port 8998. It is aimed at researchers and developers building conversational voice agents, and because it is a pretrained base model rather than a finished product, it is best suited to experimentation and further fine-tuning.