Japanese AI developer higurashi released Haru-4B on November 13, offering researchers downloadable weights for a voice model designed to listen and speak at the same time. Its model card suggests an NVIDIA GPU with about 12 GB of memory for its local example, making it a potentially interesting experiment for owners of consumer hardware.
This is a pretrained research model, not a finished, reliable voice assistant. The team warns that it is not a dependable source of facts and does not follow instructions like an assistant. AiBrainWorX has not run it or verified the hardware requirement.
How it handles conversation
Haru processes incoming user audio, outgoing model audio and model text as synchronized streams. According to its technical explanation, the system uses a Qwen3.5-4B language backbone and represents speech in frames at 7.5 per second, or roughly 133 milliseconds per frame. A small generation head produces the next audio representation in one sampling step.
The goal is to support conversational behaviors such as short acknowledgments and interruptions without treating every exchange as a separate recording followed by a response. The authors say they trained on real speech and reconstructed overlapping dialogue; those are the team’s descriptions, not independently audited claims about its training data or real-world behavior.
What is available
The model card lists nine primary languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. It says the weights use Apache 2.0, while the project repository provides code and local interfaces. Review component licenses and current documentation before redistribution or commercial deployment.
The team reports approximately 65 milliseconds in a particular int8 MacBook M1 Max test, excluding audio framing; its own write-up says framing can add up to 133 milliseconds. Those measurements do not establish latency on a Windows computer or an RTX 3060, and actual performance depends on hardware and audio setup.
Voice and safety
Haru uses a reference recording to set its voice. The publisher says it ships without voice clips and asks users to provide a voice they have the right to use. Obtain the speaker’s permission and establish how audio is stored before testing with another person’s voice.
Small local models can make voice research more accessible, but natural timing is only one part of a useful assistant. Accuracy, interruption handling, consent and language quality all require independent testing in the intended environment.
Original reporting from publisher sources; no independent model or performance test.
