# Haru-4B — an open speech model that listens and talks at the same time

> An Apache-2.0 voice model that talks like a phone call: it listens while it speaks instead of waiting for its turn.

- Org: higurashi
- Date: 2026-11-12
- Category: model
- Source: https://huggingface.co/higurashi-ai/Haru-4B
- Canonical page: https://ai-tldr.dev/releases/higurashi-haru-4b/

## In short

Haru-4B is higurashi's open full-duplex spoken dialogue model: Apache 2.0, built on Qwen3.5-4B, nine primary languages. It hears and speaks at once, with about 65 ms end-to-end latency on a MacBook.

## What it is

Haru-4B, from the Japanese company higurashi, is a full-duplex spoken dialogue model that listens and speaks at the same time, the way people do on a phone call. A text character card sets who it plays, and a short reference clip (3 seconds to 3 minutes of one speaker) sets its voice. It is a pretrained base model for research, so it is not a reliable source of facts and does not follow instructions like an assistant.

## How it works

Every 133 ms the model takes one frame of the user's audio and produces one frame of its own audio and text, so listening and speaking never wait for each other. A Qwen3.5-4B backbone reads 24 kHz audio from a VibeVoice acoustic tokenizer at 7.5 Hz, and a small MLP head with one-step MeanFlow sampling writes continuous audio latents. higurashi pretrained it on over 40 million hours of speech in 160+ languages, plus 13 million hours of separated two-speaker dialogue, on 256 B200 GPUs for about 8 days.

## Why it matters

Most open voice assistants still take turns: they wait for silence, then answer. An Apache-2.0 model that can talk over, interrupt and back-channel in nine languages gives builders a local base for natural voice agents. higurashi reports 65 ms p50 end-to-end latency on an M1 Max with int8 weights, and ships an MLX build with a native macOS app plus a vLLM-Omni plugin.

## Who it's for

researchers and developers building voice agents

## Try it

git clone https://github.com/higurashi-ai/Haru && cd Haru && pip install -e ./haru && python -m haru.server --hf-repo higurashi-ai/Haru-4B

## Links

- [Model Card](https://huggingface.co/higurashi-ai/Haru-4B)
- [Technical Blog](https://blog.higurashi.ai/melte/haru4b/)
- [GitHub](https://github.com/higurashi-ai/Haru)
- [Demo](https://huggingface.co/spaces/hugging-apps/haru-4b)
- [HF Trending](https://huggingface.co/models?sort=trending)