What higurashi released
Japanese company higurashi has put Haru-4B on Hugging Face and explained it in a technical blog post dated 13 November 2026. Haru-4B is a full-duplex spoken dialogue model, meaning it hears the user and produces its own speech at the same time. A full-duplex model can be interrupted mid-sentence and can respond while someone is still talking. Most voice agents in production today chain speech recognition, a text model and text-to-speech, with a voice activity detector guessing when the user has finished. Haru-4B replaces that chain with one network that runs continuously.
The Haru-4B model card releases the weights under Apache 2.0. Hugging Face lists about 5B parameters in BF16 safetensors even though the name says 4B, because the model is built on a Qwen3.5-4B text backbone with audio components added. The Haru GitHub repository applies Apache 2.0 to the code as well. The Open Weights tracker dates the release to 12 November but lists six languages and a non-standard licence. The model card and README both say nine languages and Apache 2.0, so we've relied on those primary sources.
How the model works
Haru-4B runs at 7.5 frames per second, so each step covers roughly 133 ms of audio. For every frame, the model adds the incoming user audio embedding to embeddings of its own previous text and audio output, and the language model processes the sum as a single position. The blog post says this summing approach, compared with interleaving the streams token by token, cuts the context length to a third while keeping the text stream intact.
Audio passes through the VibeVoice acoustic tokenizer at 24 kHz and becomes a 64-dimensional continuous latent, which removes the need for discrete codebook tokens. A four-layer MLP head, trained with flow matching and then distilled with a method called MeanFlow, generates each audio frame in one sampling step. Flow matching is a generative technique that learns to turn noise into data along a smooth path, and the distillation step means no repeated denoising passes at inference time. Text is grouped into chunks with padding positions and an end-of-padding token that signals when speech should begin.
Builders control Haru-4B with two inputs. A short text "character card" sets the persona, and a reference clip of between 3 seconds and 3 minutes from one speaker sets the voice, according to the model card.
Training data and speed
Higurashi says it pretrained on more than 40 million hours of speech covering over 160 languages, with nine primary languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. The full-duplex stage used over 13 million hours of two-channel dialogue that the team produced by separating mixed recordings into individual speakers. The company says it left out LLM-generated dialogues and synthesised speech. Training ran on 256 NVIDIA B200 GPUs for about eight days.
The headline latency figure is 65.2 ms end to end on an Apple M1 Max, using int8 weights, a native Swift app and custom Metal kernels. One model step accounts for 34.2 ms of that total. The blog post notes the 65 ms excludes audio framing, which can add up to 133 ms of algorithmic latency, so the delay a user hears can be noticeably longer. On a GPU server using vLLM-Omni with an NVIDIA B300, encoding, decoding and one model step take about 15 ms.
For comparison, Kyutai's Moshi, the best-known open full-duplex model, runs at 12.5 Hz with a theoretical latency of 160 ms and practical latency of about 200 ms on an L4 GPU, uses a 7B main transformer and releases its weights under CC-BY 4.0. Higurashi doesn't benchmark Haru-4B against Moshi, so these figures come from different hardware and different measurement methods and can't be compared directly.
Limits builders should plan around
Haru-4B is a pretrained base model released for research. The model card warns that Haru-4B isn't a reliable source of facts and doesn't follow instructions the way an assistant model does. The blog post adds that the small backbone limits response quality and that separation errors and noise remain in the training data, especially where speakers overlap.
Higurashi hasn't published any quantitative benchmark scores. The only evidence of conversational quality is a demo conversation. A detailed technical report and the training code are listed as future work, which makes fine-tuning harder today, since the fine-tuning recipe is the part most product teams would need.
Te reo Māori isn't among the nine primary languages, and nobody has published results on New Zealand-accented English. Teams building for Aotearoa will need to test both themselves before they rely on the model. Voice cloning from a short clip carries obvious consent risks, and the model card asks users to clone only voices they have the right to use.
Running Haru-4B on your own hardware
The README lists about 12 GB of GPU memory for the NVIDIA path, and a separate MLX build for Apple silicon with a native macOS app. After cloning the repository, the web demo starts with these commands:
pip install -e ./haru
python -m haru.server --hf-repo higurashi-ai/Haru-4BThe demo is served at http://localhost:8998. Browsers only allow microphone access on localhost or over HTTPS, so a remote test box needs a certificate or a tunnel. The README offers a --share flag, but that option routes audio through Gradio's share servers, which defeats the purpose of local hosting if your test recordings contain personal information. For Mac users, the README builds the app from the haru_mlx folder using ./build_app.sh in its swift subfolder. For many concurrent conversations, the README recommends the vLLM-Omni backend on a GPU server.
The README also recommends wired headphones and a wired microphone, because Bluetooth adds 100 ms or more of delay. For latency-sensitive work, that recommendation matters as much as model speed.
What a builder could do with Haru-4B
The strongest case for Haru-4B in New Zealand is local inference. A voice agent that calls a hosted speech API in the United States pays a trans-Pacific round trip on every exchange, while Haru-4B can run on a MacBook or a single workstation GPU in the same room as the user. Teams prototyping phone receptionists, kiosk assistants or language practice tools can use Haru-4B to test turn-taking behaviour, such as interruptions and short acknowledgements, without sending audio offshore.
Haru-4B is also a useful research testbed. A team could record its own interruption scenarios, measure the real latency through its own audio hardware, and compare Haru-4B against a cascaded pipeline on the same tasks. Anyone attempting a domain fine-tune should expect to write their own training loop until higurashi releases its training code. Fact-sensitive answers should be handled by a separate system, because the model card says Haru-4B isn't reliable on facts.
If you build something with Haru-4B, especially tests on New Zealand English or te reo Māori, we'd like to see it on the ASAI projects page.
Sources
- Haru-4B technical blog post (higurashi)
- higurashi-ai/Haru-4B model card, Hugging Face
- higurashi-ai/Haru GitHub repository
- Haru-4B brings full-duplex speech, The Open Weights
- kyutai-labs/moshi GitHub repository
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.