Newsletter image

Subscribe to the Newsletter

Join 10k+ people to get notified about new posts, news and tips.

Do not worry we don't spam!

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use

Search

GDPR Compliance

We use cookies to ensure you get the best experience on our website. By continuing to use our site, you accept our use of cookies, Privacy Policy, and Terms of Service.

Higurashi AI - Conversational AI, Text-to-Speech, Speech Recognition

Haru-4B

Haru-4B is the most-liked open model of the month and almost nobody has run it. Apache 2.0 full-duplex speech from a Japanese lab: three streams summed into one LM position, continuous latents with a one-step MeanFlow head, 40M hours of pretraining, and no published benchmarks at all.

License Apache 2.0
License Apache 2.0
TL;DR
  • Full duplex by summing three streams into one language-model position: user audio, model audio and model text share a slot, so listening and speaking never queue behind each other.
  • The 65 ms latency headline excludes framing. At 7.5 Hz that adds up to 133 ms more, so what a person experiences is about 130 ms on average and 198 ms at worst.
  • No benchmarks published at all: 1,756 likes against 5,314 downloads, and the authors call it a base model that does not follow instructions like an assistant.
System Requirements
RAMAbout 12 GB of GPU memory; the single bfloat16 weight file is 10.54 GB, roughly 5.27B parameters
GPUAny 12 GB NVIDIA card, or an Apple M1 Max via the bundled MLX app at int8
VRAMUse wired headphones and a wired mic: Bluetooth adds 100 ms or more to a 133 ms frame budget

Haru-4B is the most-liked open model released this month and almost nobody has run it. On the day we checked it had 1,756 likes on Hugging Face against 5,314 downloads, a ratio that inverts every other release on the board: Zyphra's ZONOS3-Mini has a third of the likes and twenty times the downloads. The reason is in the last paragraph of its own model card, where the authors say plainly that this is a pretrained base model released for research, which "does not follow instructions the way an assistant does." People are bookmarking it, not deploying it, and they are right to.

They are also right that it matters. Haru is a full-duplex spoken dialogue model: it listens and speaks at the same time, the way people do on a phone call, rather than taking turns like a walkie-talkie. It is Apache 2.0, it is 4B, and it comes from a Japanese org with three public repositories. The architecture is genuinely new, the training bill is enormous, and the headline latency figure leaves out the largest single component.

What shipped

FactValue
Published2026-11-12, weights and code the same day
LicenceApache 2.0
WeightsOne file, 10,540,113,510 bytes in bfloat16
Language modelQwen3.5-4B text backbone, 32 layers, hidden size 2560
Audio codecVibeVoice acoustic tokenizer, causal, 24 kHz
Frame rate7.5 Hz, one frame every 133.3 ms
Languages9 primary, pretrained across 160+
To run itAbout 12 GB of GPU memory
Benchmarks publishedNone
GitHub632 stars, 59 forks, 9 open issues

Note the weight file against the name. At bfloat16, 10.54 GB is roughly 5.27 billion parameters, not four billion. The "4B" refers to the Qwen3.5-4B language backbone; the audio codec, the two projection stacks and the output head account for the rest. That is honest naming by the usual standards of this field, but plan storage for 10.5 GB rather than for a 4B model.

Three streams, one position

The interesting decision is how Haru packs audio and text into a language model. Most speech-to-speech systems interleave tokens: a chunk of user audio, then a chunk of model text, then a chunk of model audio, laid end to end along the sequence. That works, and it multiplies your context length by the number of streams.

Haru sums them instead. In each 133 ms frame, the user's speech latent, the model's previous text token and the model's previous speech latent are added element-wise into a single position of the language model. The authors say this cuts context length to one third of what interleaving would need. Because all three streams occupy the same position, listening and speaking never queue behind each other, which is what "native full-duplex" actually means here rather than as marketing.

The attention stack underneath is a hybrid: of the 32 layers, 24 use linear attention and 8 use full attention, placed at every fourth layer. That is the usual trade for long sequences, and a conversation at 7.5 frames per second generates sequence length fast enough to need it.

Continuous latents, and a head that samples once

The second decision is that Haru has no discrete audio tokens at all. Audio is a 64-dimensional continuous latent per frame. A four-layer MLP head turns the language model's hidden state into that latent in one sampling step, using MeanFlow, with a guidance scale between 1 and 2 and a default of 1.5. The head was trained with flow matching first and then distilled to the single-step form.

What that buys is the absence of a second transformer. Many speech models bolt an acoustic transformer onto the language model to turn hidden states into audio tokens, and that second model runs on every frame. Haru replaces it with an MLP and one sampling step, which is the main reason a 5.27B model can hold a real-time conversation on a laptop.

65 ms, and the 133 ms that is not in it

The latency figure in the technical blog is 65 ms, measured on a MacBook with an M1 Max using int8 weights, a native Swift app and custom Metal kernels. It is a careful measurement, reported as a p50 with error bars, and broken down honestly.

Stagep50 latency
Microphone into the app5.3 ± 0.0 ms
One model step34.2 ± 0.5 ms
Socket and handoff4.1 ± 0.3 ms
Output wait, 15 ms margin and slack13.3 ± 1.5 ms
App into headphones8.1 ± 0.2 ms
Total measured65.2 ± 1.5 ms
Framing, excluded from the aboveup to 133 ms

That last row is the one to read twice, and the blog does state it: the 65 ms figure excludes audio framing, and framing adds up to 133 ms of algorithmic latency at the 7.5 Hz rate. The model cannot act on your speech until the frame containing it is complete, and where your words fall inside that 133 ms window is chance.

So the number a person actually experiences is roughly 130 ms on average and about 198 ms at worst, not 65 ms. That is still very good. Human conversational turn-taking gaps sit around 200 ms, so Haru lands inside the window where an exchange stops feeling like a radio call, which is the whole point of the architecture. It is simply not the number on the tin, and anyone budgeting a product against 65 ms will be out by a factor of two to three.

On a GPU the processing side is faster still: the authors report about 15 ms for audio encode, decode and one language model step when serving through vLLM-Omni on an NVIDIA B300. The 133 ms of framing does not move, because it is a property of the frame rate rather than of the hardware.

What it cost to train

The training scale is the part that is hard to argue with. Pretraining ran on 256 NVIDIA B200 GPUs for about eight days with Megatron-LM, over more than 40 million hours of speech spanning 160 or more languages. A second full-duplex stage used more than 13 million hours of separated dialogue, with individual samples running up to an hour.

One line in that description is worth more than the hour counts. The authors say the dialogue data excludes anything generated by a language model or synthesised by a TTS system. Training speech models on synthetic speech is common and it quietly teaches the model to imitate a synthesiser. Ruling it out is a deliberate and expensive choice, and it is the kind of methodological detail that usually only appears when someone is confident about their data pipeline.

The licence stack is clean, which is not the usual outcome

Haru is Apache 2.0, built on a Qwen backbone and a Microsoft codec, so we checked whether that claim actually survives its dependencies. It does.

ComponentSourceLicence
Language backboneQwen3.5-4BApache 2.0
Audio codecMicrosoft VibeVoiceMIT
Haru-4B weightshigurashi-aiApache 2.0

That is a permissive chain end to end, and it is worth pointing out because the last time we traced a Qwen derivative the answer was different. Edge0-145B inherited the Qwen Community License, so a commercial coding assistant needed permission from two parties rather than none. Qwen ships models under both licences, and which one you inherit depends entirely on which checkpoint you built from. Haru picked an Apache 2.0 one.

The one blemish is clerical: there is no LICENSE file in the weights repository. The Apache 2.0 declaration lives in the model card's metadata and in the GitHub repository, which is enough to be clear about intent, but a LICENSE file next to the weights costs nothing and removes the question.

There are no benchmarks

Haru publishes no evaluation numbers. Not a weak set, not a selective set: none. There is a qualitative demo conversation about investment advice, and that is the entire published evidence of how well it works. There is also no quantitative comparison to Moshi, FLM-Audio or OmniFlatten, which appear only as citations, and the promised technical report and training code have not been released.

For a model with 1,756 likes this is remarkable, and it explains the download gap better than anything else. Nobody can tell you whether Haru is better than the full-duplex models that already have weights on disk, because its authors have not said and nobody else has measured it. The architecture is novel enough to be interesting on its own terms, and the training scale is real, but "interesting architecture" and "works well" are different claims and only one of them is currently supported.

The authors are candid about the weaknesses they do know. Training used separated audio rather than native two-channel recordings, so separation errors and noise survive into the model, especially where speakers overlap. And they write that the "relatively small language-model backbone limits its response quality," which is a fair description of what a 4B conversational model can do.

It costs more memory than the pipeline it replaces

The appeal of a single full-duplex model is that it collapses three processes into one. We worked through the alternative recently: a local voice assistant built from separate parts runs about 4.4 GB with modest ears, or 5.2 GB with accurate transcription and a speculative decoding drafter.

Haru wants about 12 GB of GPU memory. The single-model route is architecturally cleaner, genuinely full-duplex in a way a pipeline can never be, and roughly two to three times the memory. That is the trade, and it is the opposite of what "one model instead of three" sounds like it should mean. If you are building for an 8 GB device, the pipeline still wins on arithmetic even though it loses on elegance.

Where it sits beside the other open full-duplex models

This is the fourth open full-duplex model we have covered, and the field has filled in quickly. We mapped it when Block announced Dessa and shipped nothing: three open models already had weights on disk while the proprietary announcement had none. PersonaPlex-7B remains the one to beat for anyone who wants voice and role control with an instruction-following model behind it.

Haru is not a competitor to those on capability, because it is a base model and they are not. It is a competitor on architecture, and specifically on the claim that continuous latents at a low frame rate beat discrete audio tokens. If the fine-tuned version the authors are planning lands, that claim gets a fair test. Today it is a very well-funded hypothesis with weights attached.

Running it

The repository ships a PyTorch implementation with a web demo, an MLX implementation with a native macOS app, and a plugin for serving through vLLM-Omni. The model has no voice of its own: you supply a reference clip, and the authors are explicit that you should only use voices you have the right to use.

git clone https://github.com/higurashi-ai/Haru && cd Haru
pip install -e ./haru

# about 12 GB of GPU memory; downloads 10.5 GB of weights on first run
python -m haru.server --hf-repo higurashi-ai/Haru-4B

# then open http://localhost:8998, pick a character, give it a voice clip

One practical note from the authors that applies to any full-duplex system: use wired headphones and a wired microphone. Bluetooth adds 100 ms or more, which is most of the latency budget this architecture exists to protect. Spending eight days on 256 B200s to save 70 ms and then routing the audio over Bluetooth is a way to measure your own headphones.

Who should care

Researchers working on spoken dialogue. This is the audience the authors named and it is the correct one. The summed three-stream representation and the single-step MeanFlow head are both reusable ideas, the licence lets you build on them, and 40 million hours of pretraining is not something most labs can reproduce.

Anyone building voice products, as a thing to watch rather than ship. The architecture is the most convincing argument yet that full-duplex does not require a second acoustic transformer. Wait for the instruction-tuned release and for somebody to publish numbers.

Not people who need an assistant today. It is a base model, it will not follow instructions reliably, it has no published evaluation, and it needs 12 GB. Every one of those is stated by its own authors, which is more than many releases manage, and all four are disqualifying for production.

Sources and further reading

The ten-minute version: read the config file before the blog post. frame_samples divided by sample_rate gives you 133 ms, and that single division tells you what the latency headline left out and what the architecture is really built around.

Not independently tested. Every figure here comes from the Haru model card and technical blog, the config file in the weights repository, the GitHub and Hugging Face APIs, or our own arithmetic on those numbers. No part of this model was run on our hardware, and the authors have published no benchmarks for us to check.
Date checked: 2026-11-19

Prev Article
Kling Image 4.0
Next Article
OpenThinker-32B

Related to this topic: