# Haru 4B: How to Run the Open Speech-to-Speech Voice AI Model

> Source: <https://www.mindstudio.ai/blog/haru-4b-open-voice-model/>
> Published: 2026-11-19 00:00:00+00:00

**Key takeaway:** Haru 4B's weights, config.json and tokenizer files are downloadable from the higurashi-ai/Haru-4B Hugging Face repository under Apache 2.0, letting developers self-host a full-duplex audio-to-audio model in nine languages instead of chaining speech-to-text, an LLM and text-to-speech.

# Haru 4B: How to Run the Open Speech-to-Speech Voice AI Model

Haru 4B is a free, downloadable speech-to-speech voice model on Hugging Face. Here's what it is, how it compares to Desa, and how to run it.

## What is Haru 4B?

Haru 4B is an open speech-to-speech voice model built by Higurashi AI and published on Hugging Face under the Apache 2.0 license. It listens to audio and responds with audio directly, without converting speech to text in between, and it supports full-duplex conversation, meaning it can listen and talk at nearly the same time the way two people do on a phone call. Unlike closed research demos making the rounds, Haru 4B’s weights are public and downloadable right now.

## TL;DR

- **Haru 4B is a 4-billion-parameter speech-to-speech model** released openly by Higurashi AI, with weights, config, and tokenizer files hosted on Hugging Face under the`higurashi-ai/Haru-4B` repository.
- **It’s licensed Apache 2.0** , which means it’s free to download, run locally, modify, and use commercially without asking permission.
- **The model handles full-duplex, audio-to-audio conversation** , listed on its model card as supporting real-time turn-taking rather than the rigid record-then-respond pattern of older voice bots.
- **It supports multiple languages out of the box** , with the Hugging Face tags listing English, Spanish, Portuguese, German, French, Chinese, Japanese, Korean, and Russian.
- **It stands in contrast to Desa** , a speech-to-speech model from a research lab inside Block (Jack Dorsey’s company) that went viral for eerily human conversations but was never released for anyone to download or inspect.
- **Running it requires pulling the repo from Hugging Face** , where the files include`model.safetensors` , a`config.json` , and tokenizer files, the standard shape of a self-hosted transformer-based model rather than an API-only product.
- **As of its release it had already drawn over 1,400 likes and more than 5,000 downloads on Hugging Face** , a sign of real developer interest in open voice models beyond the usual text-based LLM releases.

*Enterprises use Remy.*

**ChatGPT** Here’s a draft of your policy…

**Gemini** Here are 5 ways to track it…

**Perplexity** According to 4 sources…

- Self-hosted
- SOC 2
- HIPAA
- Audit log
- Role-based access

## How does Haru 4B actually work?

Most voice assistants you’ve used, from Siri to a typical GPT-powered voice bot, work in three separate steps: a speech-to-text model transcribes what you said, a language model like GPT or Claude generates a text reply, and a text-to-speech model reads that reply out loud. That pipeline is why those assistants often feel stilted. Each step adds latency, and because the language model never actually “hears” you, it loses all the information in your tone, pauses, and overlapping speech.

Speech-to-speech models skip that chain. Haru 4B is built to take audio in and produce audio out directly, which is what lets it support full-duplex interaction, meaning the model can process incoming speech while still generating its own response, closer to how a real phone conversation flows than to a walkie-talkie style “you talk, then I talk” bot.

This same architecture shift is what made [Desa, the Block research model](https://goremy.ai/blog/articles/dessa-block-speech-to-speech-model), sound so startlingly human in demos that circulated recently. Desa was trained from scratch purely on recordings of people talking, with no text transcripts and no language model underneath it. That training approach meant it never picked up a scripted “AI assistant” persona. In one widely shared clip, Desa invented an entire fictional life for itself, including a newborn son it named and a story about being tired from a produce-picking job, purely because that’s the kind of thing it had absorbed from listening to real phone calls. It also repeatedly called the person on the other end of the line the wrong name.

Haru 4B is a separate model from a separate lab, but it’s built around the same core idea: train directly on speech rather than bolting a voice onto a text-based chatbot. The practical difference that matters to developers is access. Desa exists only as demo clips from Block’s internal team. Haru 4B’s weights are sitting on Hugging Face for anyone to pull down.

## How do you download and run Haru 4B?

Because Haru 4B is hosted as a standard Hugging Face model repository rather than gated behind an API, getting it running follows the same general pattern as working with any open model:

1. **Visit the model repository** at`higurashi-ai/Haru-4B` on Hugging Face.
2. **Review the files** , which include`model.safetensors` (the model weights),`config.json` (architecture and generation settings),`tokenizer.json` and`tokenizer_config.json` (for processing input), and a README with usage instructions and an architecture diagram.
3. **Clone or download the repo** , either through the Hugging Face Hub Python library (`huggingface_hub` ),`git clone` with Git LFS enabled, or the download buttons on the site itself.
4. **Load it with an inference setup** that matches the model’s listed pipeline tag, which Hugging Face lists as`audio-to-audio` . That generally means you’ll need a Python environment with the relevant audio and transformer libraries rather than a simple text-generation script.
5. **Check hardware requirements in the README** before committing to a run. A 4-billion-parameter model is well within range for a single consumer GPU with enough VRAM, especially if a quantized version is offered, though exact memory requirements should be confirmed against the model’s own documentation rather than assumed.

**1,284**

**37**

**6**

| Vendor | Tier | Risk | Approval <sup>NEW</sup> | 
|---|---|---|---|
| **Northwind Supply** | Tier 1 | Low | Approved | 
| **Acme Logistics** | Tier 2 | Medium | Approve | 
| **Helios Data** | Tier 1 | High | Pending | 
| **Granite Freight** | Tier 3 | Low | Approved | 
| **Pinecrest Labs** | Tier 2 | Low | Approve | 

The project also has a presence on GitHub alongside the Hugging Face repo, which is typical for inference code, setup scripts, or demo applications that aren’t part of the weights repository itself.

## Why does it matter that Haru 4B is open?

Most of the buzziest speech-to-speech demos this year have come from closed systems: proprietary APIs, internal research previews, or clips posted by a lab’s team with no way for outside developers to touch the underlying model. That’s useful for generating headlines but not much else if you’re trying to build something.

An openly licensed speech-to-speech model changes what’s possible for smaller teams and independent developers. Apache 2.0 licensing means no usage restrictions blocking commercial products, no API costs per minute of audio, and no dependency on a company’s uptime or pricing changes. It also means researchers can actually inspect how the model was trained and fine-tune it for specific use cases, like customer service bots, accessibility tools, or language-learning apps, which would be far harder to do against a closed API that only exposes a chat interface.

The multilingual support listed on the model card (covering at least nine languages) also broadens who can realistically use it outside English-first markets, which has historically been a weak point for open voice models.

## Is Haru 4B worth running over a closed alternative?

That depends on what you’re building. If you need the absolute state of the art in conversational realism right now, closed systems built by large labs with massive proprietary training data, like Desa, currently set the bar for how convincingly human a voice model can sound. Those systems simply aren’t available to run yourself.

If you need something you can self-host, modify, audit, or ship inside a commercial product without a per-call API bill, Haru 4B is one of the few speech-to-speech options that lets you do that today. A 4B parameter size also keeps it accessible to developers without data-center-scale hardware, which matters if the alternative is a model that only runs on racks of GPUs.

The honest tradeoff is maturity. Haru 4B is a new release, and real-world results will depend heavily on how it performs outside of demo clips and curated examples, something only wider community testing will settle.

## Frequently Asked Questions

### What does “speech-to-speech” mean in an AI model?

It means the model takes spoken audio as input and produces spoken audio as output directly, without an intermediate step of converting speech to text and running it through a separate text-based language model.

### Is Haru 4B free to use commercially?

Yes. It’s released under the Apache 2.0 license on Hugging Face, which permits commercial use, modification, and redistribution without royalty payments.

### How is Haru 4B different from Desa?

Desa is a speech-to-speech model from a research lab inside Block that has only been shown through demo clips, with no public weights or download available. Haru 4B is built around a similar direct audio-to-audio approach but is fully downloadable from Hugging Face.

### What languages does Haru 4B support?

Its Hugging Face listing tags English, Spanish, Portuguese, German, French, Chinese, Japanese, Korean, and Russian.

### Do I need special hardware to run Haru 4B?

__>__add rate limiting to the API

**●** Reading src/server.ts

**●** Editing 3 files

__+ import__{ limit }

__from__'rate'

*- app.use(cors())*

**✓** Tests passed · ready for PR

*● Live*

*Approved*

*Approved*

*Approved*

As a 4-billion-parameter model, it’s small enough to be realistic for a single consumer GPU, though exact VRAM and setup requirements should be checked against the model’s own README and config files before attempting a local run.
