Haru 4B is a full-duplex conversation research base, not an ordinary voice generator.

The short answer

Haru 4B is described by higurashi as a low-latency, multilingual, full-duplex continuous spoken-dialogue model. “Full-duplex” is the key distinction: the system is built to process incoming speech while generating the model’s speech and text, so a conversation can include backchannels, interruptions, overlapping speech, and changing turn boundaries.

That makes Haru 4B a research base for spoken interaction, not a normal text-to-speech catalog, narration tool, or voice-cloning recommendation. A short reference clip controls the voice the model speaks with, but that does not turn the project into a conventional voice-cloning product.

The official product boundary

The official product page calls Haru 4B an “end-to-end full-duplex S2S model” and says it can listen and speak simultaneously. It contrasts this with a traditional cascaded ASR–LLM–TTS system, where speech is transcribed, answered, and synthesized in sequence. Haru instead processes audio directly and learns not only what to say, but also when and how to respond.

Most importantly, the same page says the released weights are an open-weight pre-trained base model, and explicitly notes that Haru 4B is “a pretrained base model, rather than an instruction-tuned assistant.” The model card and repository add the practical warning: it can hold a natural conversation, but it is “not a reliable source of facts” and does not follow instructions the way an assistant does. Treat those as official limitations, not as a small footnote.

In practical terms, this is closer to an open foundation for research into real-time spoken interaction than to a ready-made assistant. The character card and reference audio are conditioning inputs; they are not evidence of assistant reliability, production voice quality, or a general-purpose factual system.

What the official materials describe

The model card and repository describe three synchronized streams: user audio, model audio, and model text. They are combined at a 7.5 Hz frame rate, approximately every 133 ms. The model card describes a 64-dimensional audio latent per frame and a one-step MeanFlow audio head; the repository describes the same continuous-audio approach and identifies the VibeVoice acoustic tokenizer in the audio path.

The product page presents the design as a way to handle backchanneling, interruptions, and overlapping speech in a continuous conversation. The technical blog, dated 2026.11.13, explains the research motivation as learning conversational behavior from real-world speech rather than relying on synthetic conversations or explicitly scripted turn-taking. Those are higurashi’s technical and product descriptions, not results independently reproduced by this site.

Languages: two different official claims

The Hugging Face model card and repository list nine languages for the model: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean. That is the supported-language layer this review can safely present.

The technical blog separately says the training data spans over 160 languages, with a primary focus on those nine. Training-data span is not a verified list of product languages, so this page does not turn the broader figure into a multilingual support promise.

The product demos are vendor samples

The official product page presents four conversation recordings: a Japanese conversation about the economy, an English conversation about movies, a Japanese conversation about parenting, and an English conversation that shifts from casual conversation to a serious professional tone around financial decisions. The page asks listeners to notice turn-taking, laughter, backchannels, interruptions, overlapping speech, empathy, and tone shifts.

Those are useful examples of what higurashi wants to demonstrate, but they remain vendor samples. Voice AI Index did not listen to them as a test, score them, or compare them with another model.

Official ways to run it

NVIDIA GPU and local server

The model card and repository give a PyTorch path for an NVIDIA GPU with about 12 GB of memory. The documented server opens at http://localhost:8998. This is an official local-server route; it is not a hosted public playground and this page does not claim that every 12 GB system will provide the same experience.

Apple silicon and the native macOS app

The official MLX README describes Haru.app for Apple silicon, macOS 13 or later. It says the model is about 10 GB on disk and that 16 GB of memory is the practical minimum. The app runs the model in a Python process started by the app. Its default precision is BF16; the README also lists int8 as an alternative setting. We did not install or run the app.

vLLM-Omni

The repository includes a haru_vllm plugin for serving Haru with vLLM-Omni. That path is presented for GPU serving and concurrent conversations. It is an official code path, not a claim that this page tested throughput or deployment reliability.

These links are documentation for readers who want to inspect the project. No installation, model download, or public sharing was performed for this review.

Reference audio and voice rights

Haru ships without voice clips. The official model card and repository say to upload or record a clip for the voice the model should speak with: one speaker, 3 seconds to 3 minutes, clean speech. The official instruction is to use only voices you have the right to use.

That is a consent and rights boundary, not a promise of identity preservation or commercial permission. Apache 2.0 covers the code and Haru-4B weights; it is not a grant of rights to a person’s voice, reference recording, likeness, or generated performance. Check permission, publicity and privacy rules, contracts, platform terms, and any other applicable rights before recording, cloning, publishing, or monetizing an identifiable voice.

License and third-party conditions

The repository LICENSE and the model card identify the code and Haru-4B weights as Apache License 2.0. The Apache license includes conditions such as retaining notices and providing the license when redistributing. It does not grant a trademark license, and it does not answer the separate question of voice rights.

Haru’s official THIRD_PARTY_NOTICES.md says the project includes or derives from third-party work. It names VibeVoice, the Moshi web client, mlx-lm, and bundled web-client packages with MIT, BSD-3-Clause, ISC, and related notice conditions. Keep the required notices and read the complete file before redistributing a build.

Price and hosted access

No public price was found in the official sources reviewed here. Haru 4B is released with weights and code for research, and the Hugging Face model page says it is not deployed by an Inference Provider. This page therefore does not call Haru a free online tool, does not add an “online try it” button, and does not treat the unrelated Hugging Face Space as an official hosted Haru recommendation.

Bottom line

Haru 4B is worth examining if your question is how open models can support natural, full-duplex spoken interaction. It is not the right category for ordinary TTS shopping, a standard voice-cloning shortlist, or factual assistant selection. The useful next step is to read the official code and model documentation, check your hardware and rights, and keep the research-base limitation in view.

Official sources re-read

• Hugging Face model card
• Haru GitHub repository
• Repository LICENSE
• THIRD_PARTY_NOTICES.md
• Haru for macOS README
• Official technical blog
• Official Haru 4B product page

Method and date

This is an independent, unofficial source review. Source verification date: 2026-11-14, Asia/Shanghai. We did not install the model, run inference, record a voice, or treat the four official conversation samples as Voice AI Index listening tests. Product pages, code, demos, and runtime support can change; verify them again before a research or production decision.