What is Haru-4B?
Haru-4B is a full-duplex speech-to-speech model from higurashi, a Japanese speech AI start-up, released on 12 November 2026 under Apache 2.0. Full duplex means it listens and speaks at the same time, as people do on a phone call, rather than waiting for you to finish a turn. Every 133 milliseconds it takes one frame of your audio and produces one frame of its own audio and text; the three streams are added together into a single position of a language model built on Qwen3.5-4B. Speech is represented as continuous 64-dimensional latents rather than discrete audio tokens, and a small MLP head produces each one in a single sampling step.
It was pretrained on more than 40 million hours of real speech in over 160 languages, with nine main ones (English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean), including over 13 million hours of separated two-speaker conversation up to 60 minutes long. A text character card sets the persona and a 3-second to 3-minute clip sets the voice. It is released as a pretrained base model for research, not as an assistant.
Is Haru-4B worth it?
Haru-4B is a serious open attempt at the hardest part of voice AI, talking over and around a person the way people do, and the release gives you the architecture and latency figures but no quality scores. higurashi's technical post describes backchannels, interruptions and overlapping speech learned from real recordings, and it deliberately excluded LLM-written dialogues and text-to-speech audio from training. It publishes no benchmark tables against Moshi, MiniCPM-o or any other duplex model, so how good the conversation is rests on demos.
The engineering is the strongest part. On an M1 Max MacBook, the native Swift app reports about 65 milliseconds end to end with int8 weights, of which one model step takes 34.2ms; the 7.5Hz frame rate can add up to 133ms more. On an NVIDIA B300, audio encoding, decoding and one language-model step take about 15ms, and the vLLM-Omni plugin serves many conversations at once. Running at 7.5 steps a second with continuous latents is what makes a 4B model fast enough for this on a laptop.
The scale of the pretraining is unusual for an open release: 40 million hours of speech, trained on 256 NVIDIA B200 GPUs for about 8 days, with support from AWS and Japan's GENIAC programme. Interest has been quick, at 842 likes on Hugging Face and 716 GitHub stars within days.
The catch: it is a base model. The card says it holds a natural conversation but is not a reliable source of facts and does not follow instructions as an assistant does, and the company says its small backbone limits response quality. It clones any voice from a 3-second clip with no watermark mentioned, leaving consent to the user, and the demo's `--share` flag opens a public link that routes audio through Gradio's servers.
Our take: for a lab or product team building a voice companion, Haru-4B is the open checkpoint to fine-tune on its own two-channel recordings; for a support line that must give correct answers, it is a foundation to study, not something to put on the phone.
Where it earns its place
A speech research group at a university
Study turn-taking and overlapping speech in an open model
Run the PyTorch implementation on one 12GB NVIDIA GPU and use offline inference to feed recorded user audio frame by frame. Weights, code and architecture are open; a fuller technical report and training code are promised but not yet out.
A start-up building a language-practice voice companion
Give learners a partner that reacts mid-sentence instead of waiting
Write a character card per tutor and supply a licensed voice clip. Expect to fine-tune before launch, since the base model does not follow instructions; serve production traffic through the vLLM-Omni plugin.
A Mac developer prototyping an on-device voice app
Test low-latency duplex speech without a cloud bill
Build the native Haru.app from the MLX implementation with `./build_app.sh` and use wired headphones; the company reports about 65ms latency on an M1 Max.
Before you commit
- Setup
- `git clone https://github.com/higurashi-ai/Haru`, `pip install -e ./haru`, then `python -m haru.server --hf-repo higurashi-ai/Haru-4B` and open localhost:8998; you must upload your own voice clip, since none ship with it.
- Integrations
- PyTorch with a React web client, MLX and a native Swift macOS app, a vLLM-Omni serving plugin, and an offline `haru-inference` command; Triton codec kernels for NVIDIA GPUs.
- Data and privacy
- Runs entirely on your hardware unless you use `--share`, which tunnels audio through Gradio's share servers to anyone holding the link. Trained on real speech cleaned with Sidon speech restoration and split into two channels with the company's FastDialogueSidon model; sources are not itemised.
- Cost at scale
- No licence fee. One conversation fits on a 12GB NVIDIA GPU; for many concurrent calls the vLLM-Omni path targets data-centre GPUs, where a B300 runs a full step in about 15ms. Per-GPU concurrency is not published.
- Limits
- 24kHz audio in and out, one 133ms frame at a time; reference voice 3 seconds to 3 minutes of one speaker; no benchmarks, no instruction following, and a browser only allows the microphone on localhost or HTTPS.
What’s good
- Listens and speaks at the same time, with backchannels and interruptions
- About 65ms end-to-end latency on an M1 Max MacBook, per the company
- Apache 2.0 for both weights and code
- Nine main languages from 40 million hours of real speech
- PyTorch, MLX and vLLM-Omni implementations ship together
What’s not
- No published benchmarks or comparisons with other duplex models
- Base model: unreliable on facts and does not follow instructions
- Clones a voice from a 3-second clip with no watermark mentioned
- Bluetooth audio adds 100ms or more, so wired headsets are advised
Best for
Speech researchers and voice-product teams who want an open full-duplex base model to fine-tune for natural, interruptible conversation.
Haru-4B alternatives
Haru-4B: common questions
- Is Haru-4B worth it?
- Haru-4B is an ambitious open start on full-duplex voice with laptop-class latency, but it is a research base model that clones voices and publishes no scores. We rate it 3.6 out of 5.
- How much does Haru-4B cost?
- Free weights and code under Apache 2.0. Self-hosting needs an NVIDIA GPU with about 12GB of memory or an Apple silicon Mac; no hosted API.
- What are the downsides of Haru-4B?
- No published benchmarks or comparisons with other duplex models; base model: unreliable on facts and does not follow instructions; clones a voice from a 3-second clip with no watermark mentioned; bluetooth audio adds 100ms or more, so wired headsets are advised.
- Who is Haru-4B best for?
- Speech researchers and voice-product teams who want an open full-duplex base model to fine-tune for natural, interruptible conversation.
- How much does Haru-4B cost for a team or in production?
- No licence fee. One conversation fits on a 12GB NVIDIA GPU; for many concurrent calls the vLLM-Omni path targets data-centre GPUs, where a B300 runs a full step in about 15ms. Per-GPU concurrency is not published.
- Is Haru-4B safe for company data?
- Runs entirely on your hardware unless you use `--share`, which tunnels audio through Gradio's share servers to anyone holding the link. Trained on real speech cleaned with Sidon speech restoration and split into two channels with the company's FastDialogueSidon model; sources are not itemised.
- What are the best Haru-4B alternatives?
- The closest alternatives we have reviewed are MiniCPM-o 4.5, GPT-Live-1 (and GPT-Realtime-2.1) and Sesame CSM-1B.
Answers checked Nov 21, 2026.