Haru-4B — full-duplex speech in one file, no Python

Haru-4B by higurashi-ai listens and speaks at the same time, like a person on a phone call: every 133 ms it takes one frame of the user's audio and produces one frame of its own audio and text. This repository packages it as one CMF file for cortiq, a Rust engine with no Python and no ML framework — the Qwen3.5-4B language model, the audio projectors, the MeanFlow head and the VibeVoice codec together.

  • RTX A5000, CUDA: 23 ms per 133 ms frame — 5.8× faster than real time. Vulkan 53 ms, Apple M4 (Metal) 84 ms.
  • The codec matches the PyTorch reference within 1.3·10⁻³.

Listen: a conversation — the user on the left channel, Haru on the right. The user asks "Hi! How are you doing today? Could you tell me something interesting about the ocean?"; Haru, speaking with the reference voice, answers "Hey, welcome to our daily talk show… Today I'm gonna be talking about science, but not just any science, the science of the ocean." Both input voices were made with cortiq's Qwen3-TTS.

Quick start

cargo install cortiq-cli --features cuda   # NVIDIA: the CUDA backend; without the feature: Vulkan, or Metal on a Mac
hf download infosave/Haru-4B-cmf haru-4b-q8_2f.cmf --local-dir .
CMF_GPU=cuda cortiq duplex haru-4b-q8_2f.cmf user.wav reply.wav --reference voice.wav --stereo --transcript reply.jsonl

user.wav is played to the model frame by frame, exactly as in a live call, and reply.wav is what the model says over the same span (--stereo: the user left, the model right). Leave silence after a question so the model has time to answer. --reference sets the voice: 3 s to 3 min of one speaker, clean speech — use only voices you have the right to use. Other options: --system-prompt (a character card, as the Haru repository describes it), --speak-first, --temperature (0.7), --top-k (25), --guidance-scale (1.5), --seed.

File

file size
haru-4b-q8_2f.cmf 6.37 GB the language model in 8-bit (q8_2f), the head and the codec in f16

Four-bit weights break this model: with q4tp, even with only the feed-forward layers at four bits, the text stream turns to noise. There is no smaller file.

Speed

Per 133 ms frame, after the prompt and the voice reference, which run once (4–13 s):

  • RTX A5000 (24 GB), CUDA: 23 ms.
  • RTX A5000, Vulkan: 53 ms.
  • Apple M4 (24 GB), Metal: 84 ms.

Notes

Haru-4B is a pretrained base model released for research: it holds a natural conversation, but it is not a source of facts and does not follow instructions like an assistant. It can imitate the voice of a reference clip; only use voices you have the right to use. The live, microphone-in-the-loop demo is in the Haru repository; cortiq duplex runs the same loop over a recording.

License

Apache-2.0, as the original. The weights are higurashi-ai/Haru-4B, converted to CMF with the language model re-quantized to 8 bits. This repository is not made or endorsed by higurashi-ai.


На русском

Haru-4B от higurashi-ai слушает и говорит одновременно, как человек в телефонном разговоре: каждые 133 мс берёт кадр звука пользователя и выдаёт кадр своего звука и текста. Здесь модель упакована в один файл CMF для cortiq — движка на Rust без Python: языковая модель Qwen3.5-4B, проекторы звука, голова MeanFlow и кодек VibeVoice вместе.

  • RTX A5000, CUDA: 23 мс на кадр 133 мс — в 5,8 раза быстрее реального времени. Vulkan — 53 мс, Apple M4 (Metal) — 84 мс.
  • Кодек совпадает с эталоном PyTorch с точностью 1,3·10⁻³.
cargo install cortiq-cli --features cuda
hf download infosave/Haru-4B-cmf haru-4b-q8_2f.cmf --local-dir .
CMF_GPU=cuda cortiq duplex haru-4b-q8_2f.cmf user.wav reply.wav --reference voice.wav --stereo

user.wav проигрывается модели кадр за кадром, как в живом звонке; reply.wav — то, что она говорит за то же время. После вопроса оставьте тишину, чтобы модель успела ответить. --reference задаёт голос (3 с – 3 мин одного говорящего); используйте только голоса, на которые у вас есть права. Файл один — haru-4b-q8_2f.cmf (6,37 ГБ): 4-битные веса эту модель ломают. Это базовая исследовательская модель: она ведёт разговор, но не источник фактов. Лицензия Apache-2.0, как у оригинала; репозиторий не сделан и не одобрен higurashi-ai.

中文说明

higurashi-ai 的 Haru-4B 能像打电话一样同时听和说:每 133 毫秒接收一帧 用户音频,并输出一帧自己的音频和文本。本仓库把它打包为一个 CMF 文件,由 cortiq(不依赖 Python 的 Rust 引擎)运行:Qwen3.5-4B 语言模型、音频投影器、MeanFlow 头和 VibeVoice 编解码器合为一体。

  • RTX A5000,CUDA:每 133 毫秒帧 23 毫秒——比实时快 5.8 倍。Vulkan 53 毫秒,Apple M4(Metal)84 毫秒。
  • 编解码器与 PyTorch 参考实现误差在 1.3·10⁻³ 以内。
cargo install cortiq-cli --features cuda
hf download infosave/Haru-4B-cmf haru-4b-q8_2f.cmf --local-dir .
CMF_GPU=cuda cortiq duplex haru-4b-q8_2f.cmf user.wav reply.wav --reference voice.wav --stereo

user.wav 会像实时通话一样逐帧播放给模型,reply.wav 是模型在同一时间段内说的话。提问后请留出静音,让模型有时间 回答。--reference 设定声音(同一说话人 3 秒至 3 分钟);只使用你有权使用的声音。只有一个文件 haru-4b-q8_2f.cmf (6.37 GB):4 位权重会破坏该模型。这是用于研究的预训练基础模型,能自然对话,但不是事实来源。许可证为 Apache-2.0, 与原模型相同;本仓库并非由 higurashi-ai 制作或认可。

Downloads last month
199
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for infosave/Haru-4B-cmf

Quantized
(6)
this model