Haru-4B: Talk to a Free AI That Listens and Speaks at the Same Time — Full Setup Guide

Woman talking on a mobile phone, smiling

Every voice AI you have used works the same way: you talk, it waits, it thinks, then it answers. Real conversations do not work like that — people interrupt, talk over each other, say "mm-hmm" while the other person is still speaking. Haru-4B is a free, open-source model that does it the human way: it listens and speaks at the same time, like a phone call. It was released on 12 November 2026 by higurashi-ai, it is Apache 2.0 licensed, and you can run it yourself — on a GPU, on a Mac, or free on Google Colab.

What Haru-4B is

  • Full-duplex speech. Haru listens and speaks simultaneously. Every 133 milliseconds it takes one frame of your audio and produces one frame of its own speech and text — listening and speaking never wait for each other. It handles backchannels ("uh-huh"), interruptions and overlapping speech, because it learned from real conversations, not scripted turn-taking.
  • Fast. About 65 milliseconds end-to-end latency on a MacBook, according to the developers. That is close to the delay on a normal phone call.
  • Multilingual. Nine primary languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. Training covered more than 40 million hours of speech across 160+ languages.
  • Open and free. The model weights are on Hugging Face and the code is on GitHub under Apache 2.0. The repository had 474 stars within four days of release. There is no subscription, no per-minute charge, and no account needed to download it.
  • Built on Qwen3.5-4B. Haru-4B starts from the Qwen3.5-4B language model and adds a one-step audio head, so it generates speech directly instead of bolting a text-to-speech system onto a chatbot.

What you need — the honest version

  • GPU route: an NVIDIA GPU with about 12 GB of memory. That is the project's own figure. A laptop RTX 3060 or better qualifies.
  • Mac route: a Mac with Apple silicon, using the project's MLX implementation, which also ships as a native macOS app. The checkpoint is about 10 GB and downloads from Hugging Face on first use.
  • No GPU at all: Google Colab's free tier, which gives you a T4 GPU with about 15 GB of memory — enough for Haru-4B. This is the free route below.
  • Audio gear: wired headphones and a wired microphone. The developers recommend them for low latency — Bluetooth adds 100 ms or more. And a voice clip: Haru ships without any voice, so you upload or record a 3-second to 3-minute clip of one speaker's clean speech, and the model speaks in that voice. Only use voices you have the right to use.

Step 1: Try the demos first (no install)

Before installing anything, listen to what Haru actually sounds like. The project's technical blog at blog.higurashi.ai has recorded demo conversations — including a financial-advice call in English — with video and captions. If the natural interruptions and back-and-forth in those recordings are what you want, install it. Note honestly: there is no official hosted demo or Hugging Face Space to try in your browser — the free route is running it yourself, which is what the rest of this guide covers.

Step 2: Install Haru on your own GPU

  1. Go to the official repository at github.com/higurashi-ai/Haru — this GitHub page and the Hugging Face model page at huggingface.co/higurashi-ai/Haru-4B are the only official sources.
  2. Clone the repo and install the package:
    git clone https://github.com/higurashi-ai/Haru && cd Haru
    pip install -e ./haru
  3. Start the web demo server:
    python -m haru.server --hf-repo higurashi-ai/Haru-4B
  4. Open http://localhost:8998 in your browser, pick a character, give the model a voice, and talk. A browser only gives a page microphone access on localhost or over HTTPS, which is why the local address works.
  5. To let someone else try it, add --share (install with pip install -e "./haru[share]" first): the server prints a public HTTPS link that tunnels to it, and anyone with the link can use the demo.

Step 3: The free route — Google Colab, no GPU needed

  1. Open colab.research.google.com and create a new notebook. In Runtime → Change runtime type, set the hardware accelerator to T4 GPU — the free tier's GPU has about 15 GB of memory, which covers Haru-4B's roughly 12 GB requirement.
  2. In a notebook cell, run the same commands as Step 2:
    !git clone https://github.com/higurashi-ai/Haru && cd Haru
    !pip install -e ./haru
  3. Start the server with the share flag so you get a public HTTPS link (the microphone needs HTTPS):
    !python -m haru.server --hf-repo higurashi-ai/Haru-4B --share
    The server prints a public link. Open it, allow microphone access, pick a character and a voice, and talk.
  4. Heads-up: Colab's free tier disconnects idle sessions and has daily usage limits, so this is a free way to try Haru, not a 24/7 hosted service. Keep your voice clip and character card saved so you can restart in a minute.

Step 4: Give it a character and a voice

Haru reads its character from a "card" — a short text profile with a persona, name and situation. The README shows the pattern: for example, "Haru is a 29-year-old nurse in Denver. She is warm, chatty and laughs easily", talking to an old friend from high school. Write your own card in plain words, then record or upload your voice clip — one speaker, 3 seconds to 3 minutes, clean speech. In the demo you can also save a clip as haru/haru/assets/voices/en.flac (or another language code) to make it the default voice for that language.

Step 5: Use it from Python, or offline

If you want Haru inside your own project instead of the web demo, the Python API is a few lines:

from haru import DuplexSession, HaruModel
from haru.audio import load_audio, prepare_reference

model = HaruModel.from_pretrained("higurashi-ai/Haru-4B")
voice = prepare_reference(load_audio("voice.wav"))
session = DuplexSession(model, system_prompt=card, reference=voice)
frame = session.start()
while not frame.done:
    frame = session.step(user_audio)

For a one-off recording without a live conversation, the offline command turns an input recording into a reply recording:

haru-inference user.wav reply.wav --reference voice.wav --system-prompt "$(cat card.txt)" --stereo

For many simultaneous conversations on a GPU server, the repository also includes a plugin that serves Haru with vLLM-Omni — that is the production route, not the beginner one.

What it cannot do

Be straight about the limits, because the developers are. Haru-4B is a pretrained base model, released for research. It holds a natural conversation, but it is not a reliable source of facts and does not follow instructions the way an assistant does — do not use it as a search engine or a homework machine. It can imitate the voice of a reference clip, so only ever use voices you have the right to use. And because it is a base model, its answers can wander; it is a conversation partner, not a finished product.

Bottom line

Kokoro gave you free voiceovers, ZONOS3 gave you free voice cloning — Haru-4B gives you the third piece: a free voice you can actually talk with, that listens while it speaks. A 12 GB GPU, a Mac, or a free Colab notebook is all it takes, and the whole stack is Apache 2.0 open source. If you have ever wanted a voice AI that behaves like a phone call instead of a walkie-talkie, this is the weekend project to try first.

Sources: the Haru GitHub README and Hugging Face model page, checked live 16 November 2026 (repository created 12 November 2026, Apache 2.0, 474 stars), and the higurashi technical blog. Hero photo: woman talking on a mobile phone, by Shixart1985, CC BY 2.0, via Wikimedia Commons.

Tags

Post a Comment

0 Comments
* Please Don't Spam Here. All the Comments are Reviewed by Admin.