# Haru-4B: the open AI voice model that talks and listens at once

Haru-4B from Japan's higurashi is a free, open voice model that listens and speaks at the same time, in 9 languages. What it does and how to run it.

URL: https://aitoolsthisweek.com/haru-4b
Published: 2026-11-16T03:28:41.933Z
Updated: 2026-11-16T03:28:41.933Z

**Haru-4B is a free, open voice AI model that listens and speaks at the same time, so a conversation with it feels like a phone call rather than a walkie-talkie.** Japanese company higurashi released it on November 12 under the Apache 2.0 license, and it runs on an NVIDIA GPU with about 12 GB of memory or as a native app on Apple silicon Macs.

It's been climbing Hugging Face's trending list this week. Here's what's new about it and what it can and can't do.

## What makes Haru-4B different

Most voice assistants take turns: you talk, they wait for silence, then they answer. Haru is **full-duplex**. From the [model card](https://huggingface.co/higurashi-ai/Haru-4B):

- **Listens while it talks.** Every 133 milliseconds it takes one slice of your audio and produces one slice of its own audio and text, so it can react mid-sentence.
- **Fast on a laptop.** higurashi's [technical blog](https://blog.higurashi.ai/melte/haru4b/) measured about 65 ms per step on a MacBook with an M1 Max (int8 weights), plus up to 133 ms of built-in framing delay.
- **Nine languages:** English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean.
- **You set the character and the voice.** A text "character card" sets who it is (name, persona, situation). A short clip, 3 seconds to 3 minutes of one speaker, sets how it sounds.

## How it was built

| | Haru-4B |
|---|---|
| Base language model | Qwen3.5-4B (32 layers) |
| Audio codec | VibeVoice acoustic tokenizer, 24 kHz |
| Training audio | 40M+ hours of speech in 160+ languages |
| Dialogue training | 13M+ hours of separated two-person conversations, up to 60 minutes long |
| Training compute | 256 NVIDIA B200 GPUs for about 8 days |
| License | Apache 2.0 |

The blog says LLM- and TTS-generated conversations were left out of the training data, so it learned from real human dialogue. Sources: [model card](https://huggingface.co/higurashi-ai/Haru-4B) and [blog](https://blog.higurashi.ai/melte/haru4b/).

## How to try it

From the [GitHub repository](https://github.com/higurashi-ai/Haru):

1. **NVIDIA GPU (about 12 GB):** clone the repo, run `pip install -e ./haru`, then `python -m haru.server --hf-repo higurashi-ai/Haru-4B` and open `http://localhost:8998` in your browser.
2. **Pick a character and give it a voice.** Haru ships with no voice clips, so upload or record one.
3. **Mac with Apple silicon:** the repo has an MLX version with a native macOS app you build with a script (`build_app.sh`), then open `Haru.app`.
4. **Use wired headphones and a wired mic.** Bluetooth adds 100 ms or more of lag.

Running models at home keeps your audio on your own machine; our guide to [on-device vs cloud AI](/on-device-vs-cloud-ai) explains why that matters.

## Limits to know

- **It's a base model for research.** higurashi says it holds a natural conversation but isn't a reliable source of facts and doesn't follow instructions the way an assistant does.
- **Voice imitation.** It copies the voice in your reference clip. Only use voices you have the right to use.
- **Small brain.** The blog says the 4B backbone limits answer quality, and the team plans to scale to 100B+ parameters. A detailed technical report and training code aren't out yet.

If you want a text-to-speech model rather than a conversation partner, see [ZONOS3-Mini](/zonos3-mini). Haru is built on a small Qwen model; for the bigger open Qwen, see [Qwen3.8-27B](/qwen3-8-27b).

*AI Tools This Week is not affiliated with higurashi.*

## Frequently asked questions

### Is Haru-4B free?

Yes. The weights are on Hugging Face under the Apache 2.0 license and the code is on GitHub. You need your own hardware: an NVIDIA GPU with about 12 GB of memory, or a Mac with Apple silicon for the native app.

### What does full-duplex mean?

The model hears you and talks at the same time. Every 133 milliseconds it takes in a slice of your audio and produces a slice of its own, so it can react while you're still talking, the way people do on a phone call, instead of waiting for you to stop.

### Can Haru-4B clone a voice?

It imitates the voice in a reference clip you give it (one speaker, 3 seconds to 3 minutes of clean speech). It ships with no voices, and higurashi says to only use voices you have the right to use.

## Sources

- [higurashi-ai/Haru-4B model card (Hugging Face)](https://huggingface.co/higurashi-ai/Haru-4B)
- [higurashi-ai/Haru code repository (GitHub)](https://github.com/higurashi-ai/Haru)
- [Haru-4B technical blog post (higurashi, Nov 13, 2026)](https://blog.higurashi.ai/melte/haru4b/)
