Sign In

Haru-4B Real-Time Voice Agent Studio — Full-Duplex Setup

Download

1 variant available

Archive Other

Haru4B_Realtime_Voice_Agent_Studio_Pack_Civitai_Free.zip

4.72 KB

Verified:

Type
Workflows
Stats

12

Reviews
Published

Nov 19, 2026

Base Model

Other

Hash
AutoV27F6DE84CD5
cover.jpg

https://whop.com/voicedesk/free-starter-haru-4b-real-time-voice-agent-studio

> **FULL STUDIO ($50):** https://whop.com/voicedesk/haru-4b-real-time-voice-agent-studio-full-duplex-setup — use code **HARU30** for 30% off ($50 → $35.00).

## FREE vs FULL

| | FREE (this download) | FULL ($50) |

|---|---|---|

| Setup quickstart | Yes (lite) | Complete guide (GPU + Mac + Colab + Python API + offline + vLLM-Omni) |

| Character cards | 1 (receptionist) | 12 cards |

| Demo scripts | 1 (receptionist) | 4 full scripts (receptionist, creator intro, FAQ voice, multilingual) |

| Install configs | — | 5 (CUDA / MLX / Colab + server + card schema) |

| Workflow doctor | — | Yes (7-row table) |

| Guides (EN + AR) | Quick README only | Yes |

## What this is

Haru-4B is the first open full-duplex spoken-dialogue model: it listens and speaks **at the same time**, the way people do on a phone call — no ASR → LLM → TTS stack, no turn-taking. Every 133 ms (7.5 Hz) it takes one frame of your audio and produces one frame of its own audio and text. A text character card sets the character; a short reference clip (3 seconds to 3 minutes) sets the voice. Nine languages, Qwen3.5-4B backbone, runs on ~12 GB of VRAM or on a Mac (MLX), Apache 2.0. It was the #1 trending model on Hugging Face on 19 Nov 2026 (2,136 likes, 6,266 downloads; repo added 12 Nov 2026; 639 GitHub stars at build time). This free lite kit gives you the install quickstart (GPU, Mac, Colab), one character card (receptionist) and one demo script so you can hold your first live call today. The full pack adds 11 more character cards, 3 more full demo scripts, 5 install/server configs and full EN + AR guides. Caveats: Haru-4B is a pretrained **base model, not a ready assistant** — it is not a reliable source of facts and does not follow instructions like an assistant — and voice cloning is **consent-only**: only use voices you have the right to use.

## Files you need (not included)

No model files needed. Haru-4B itself is open (Apache 2.0) — download the weights from https://huggingface.co/higurashi-ai/Haru-4B and clone the code from https://github.com/higurashi-ai/Haru. You add your own voice (your own, or one with written permission). No keys included.

Built from public research; all content original. No brand assets, no credentials. Figures cited are from the sources in ADAPTED_NOTES.md — they describe the market, not a promise.

**SFW only. 'Depicts an actual person' = No. No 18+ attestation.**