Home » Haru‑4B brings low‑latency full‑duplex voice chat to developers as open‑source model

Haru‑4B brings low‑latency full‑duplex voice chat to developers as open‑source model

By

–

Haru‑4B brings low‑latency full‑duplex voice chat to developers as open‑source model

A Japanese student team called Higurashi has released Haru‑4B, a 4‑billion‑parameter model that can listen and speak at the same time with sub‑100‑millisecond latency. The shift is real‑time, full‑duplex voice chat moving out of the lab and onto developers’ machines.

What Haru‑4B Is

Haru‑4B is a low‑latency, multilingual full‑duplex continuous spoken‑dialogue foundation model. In plain terms, it handles a phone‑call‑style conversation: the user talks while the model talks, and neither side waits for the other to finish. The model is built on top of Qwen3.5‑4B and was pre‑trained on more than 40 million hours of multilingual audio, with AI‑generated audio and dialogue deliberately excluded. A custom pipeline called FastDialogueSidon produced a full‑duplex dialogue dataset of over 13 million hours for the final training run.

How It Works

The model runs at 7.5 Hz, meaning it consumes one frame of user audio and produces one frame of its own audio plus text every 133 milliseconds. User audio, model audio and model text are summed into a single language‑model position, so listening and speaking never block each other. Audio is represented as a 64‑dimensional latent per frame and generated in one step by a small MLP head using MeanFlow sampling — no separate audio transformer, no discrete audio tokens. Character is set through a text card, and voice is cloned from a short reference clip of three seconds to three minutes of clean speech from a single speaker.

Languages and Hardware

Haru‑4B supports nine languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. On a MacBook with an M1 Max chip the model runs at roughly 65 ms latency; on an NVIDIA B300 GPU latency drops to about 15 ms. A local deployment needs an NVIDIA GPU with around 12 GB of memory and the provided Python server script.

Open‑Source Release and Licensing

Weights and code are public on Hugging Face and GitHub under the Apache License 2.0 for research use. The repository includes a PyTorch implementation with a web demo, an MLX build with a native macOS app, and a vLLM‑Omni plugin. Developers can clone the repo, launch the server and open http://localhost:8998 to pick a character, upload a voice clip and start talking. Wired headphones and a wired microphone are recommended; Bluetooth adds 100 ms or more of latency.

Phone Demo and Custom Voices

A live phone demo is available in Japanese: dialing 050‑5444‑3154 connects to a male voice, 050‑5444‑3155 to a female voice. Anyone can give Haru‑4B a custom voice by uploading a clean speech clip of three seconds to three minutes, provided they hold the rights to that audio.

Limitations

Haru‑4B is a pretrained base model. It holds a natural conversation but is not a reliable factual source and does not follow instructions the way a typical assistant does. The team notes that Higurashi has already begun pre‑training a larger full‑duplex model with more than 100 billion parameters. Watch this space.

Sources

Photo: Douglas Lopes / Unsplash (Unsplash License)

Get breaking stories first on our Telegram channel.

More stories by

Most read

Crime & Fraud

People in the News

News

World News