Skip to content
Audio

Haru-4B

Multilingual full-duplex voice dialogue model

2 min readupdated

Haru-4B is a low-latency pretrained base model designed for continuous full-duplex live conversation, enabling it to listen and speak simultaneously like a phone call. Built on a Qwen3.5-4B text backbone with approximately 5.27 billion parameters, it supports nine languages, including Russian. It is ideal for researchers and developers building real-time interactive voice assistants, bots, and applications.

Key features

  • Native full-duplex at 7.5 Hz. Every 133 ms, the model receives an audio frame from the user and generates its own audio and text frame without mutual waiting.
  • Continuous audio in a single sampling step. Audio is represented as a 64-dimensional latent vector per frame, generated by a small MLP (multilayer perceptron, a feedforward neural network) head with MeanFlow sampling.
  • Character and voice control. A text prompt defines the character's persona, while a short reference audio clip determines the desired voice.
  • Multilingual support. Full support for English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean.

What you can do

  • Launch a local web server. Set up a demo on an NVIDIA GPU using the official PyTorch repository for real-time conversation with the model.
  • Build a macOS app. Use the MLX implementation for native performance on Apple Silicon Macs.
  • Integrate into server solutions. Configure a service plugin using vLLM-Omni for rapid deployment.
  • Clone voices. Assign a custom voice to the model using a clean reference audio clip ranging from 3 seconds to 3 minutes in length.

Example

Running the server to converse with the model in the project repository:

git clone [https://github.com/higurashi-ai/Haru](https://github.com/higurashi-ai/Haru) && cd Haru
pip install -e ./haru
python -m haru.server --hf-repo higurashi-ai/Haru-4B
Read moreDetails
Backbone: Qwen3.5-4B text model, 32 layers, hidden state size 2560
Audio: VibeVoice acoustic tokenizer (24 kHz, 3200 samples per latent)
Audio head: 4-layer MLP with MeanFlow (1 sampling step)
Requirements: NVIDIA GPU with approximately 12 GB VRAM
License: Apache License 2.0
Discuss Telegram 0 21

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 9.8 GB · ready files11 GB
  • FP8 / INT8recommendedweights 4.9 GB5.9 GB
  • 4-bitweights 2.8 GB3.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 5070

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # nesoai-qa-org/higurashi-ai-Haru-4B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download nesoai-qa-org/higurashi-ai-Haru-4B --local-dir ~/models/higurashi-ai-Haru-4B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to higurashi-ai-Haru-4B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of higurashi-ai-Haru-4B (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

nesoai-qa-org/higurashi-ai-Haru-4B
Author
nesoai-qa-org
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments