This is a simplified guide to an AI model called Haru-4b maintained by Higurashi-Ai. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.
Overview
Haru-4B is a research-release pretrained base model for continuous, multilingual, full-duplex spoken dialogue: it listens and generates speech and text concurrently. It operates at 7.5 frames per second, with each frame representing about 133 ms. Maintained by higurashi-ai.
Best use cases
Use it to prototype real-time spoken conversations where the model should listen and speak at the same time, with a chosen character and a voice supplied by a reference clip. The listed languages are English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean. It is released for research, not as a dependable factual or instruction-following assistant.
Limitations
Haru-4B is a pretrained base model: its documentation says it can converse naturally but is not a reliable source of facts and does not follow instructions like an assistant. It ships without voice clips, and voice imitation should only use voices you have the right to use. The supplied sources provide no benchmark results or detailed evaluation data.
How it compares
The supplied alternatives target different tasks, so they are not direct substitutes for Haru-4B's general full-duplex dialogue. Hibiki-Zero is described as simultaneous speech translation, with French, Spanish, Portuguese, and German translated to English. Hibiki is described as streaming French-to-English speech translation. T5Gemma-TTS is a multilingual text-to-speech model, while Irodori-TTS is Japanese text-to-speech with voice and style controls. Higgs TTS 3 is described as conversational speech generation from model responses. These descriptions do not establish comparative quality, speed, or cost.
Technical specifications
The language backbone is Qwen3.5-4B, with 32 layers and hidden size 2560. A causal VibeVoice acoustic tokenizer operates at 24 kHz, with 3200 samples per latent; audio is represented as a 64-dimensional latent per frame. A 4-layer MLP MeanFlow audio head generates audio in one sampling step, with guidance scale from 1 to 2 and default 1.5. At 7.5 Hz, each frame combines user audio with the model's previous text token and speech latent at one language-model position, predicting the next text token and speech latent. Weights are in model.safetensors in bfloat16; the repository also includes config.json and tokenizer.json. The README estimates about 12 GB of memory for an NVIDIA GPU setup. The weights use Apache License 2.0.
Model inputs and outputs
Inputs
A duplex session takes user audio in 3200-sample frames at 24 kHz and uses a character card as its system prompt. Voice is set separately with a clean reference clip from one speaker, 3 seconds to 3 minutes long. Haru does not ship with voice clips. The documented Python interface prepares the reference audio with prepare_reference.
Outputs
Each frame provides generated audio and text; the documented session example returns 3200 audio samples at 24 kHz per step. The model generates continuous audio latents rather than discrete audio tokens, and produces speech and text while receiving user audio.
Getting started
git clone https://github.com/higurashi-ai/Haru && cd Haru pip install -e ./haru python -m haru.server --hf-repo higurashi-ai/Haru-4B The documented setup uses an NVIDIA GPU with about 12 GB of memory. After starting the server, open http://localhost:8998, choose a character, provide a voice reference, and talk. The repository also provides PyTorch, MLX, and vLLM-Omni serving options.
Frequently asked questions
Q: Is Haru-4B an instruction-tuned assistant?
A: No. It is a pretrained base model for research and is not documented as following instructions like an assistant.
Q: Can Haru-4B listen and speak at the same time?
A: Yes. It is designed for full-duplex dialogue, processing and generating audio and text concurrently at 7.5 frames per second.
Q: Which languages are listed?
A: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean.
Q: What voice reference should I provide?
A: Use clean speech from one speaker, between 3 seconds and 3 minutes long. The model does not include voice clips; use only voices you have the right to use.
Q: What audio format does the documented session use?
A: The example uses 3200 samples at 24 kHz for each input and output frame.
Q: What hardware does the setup documentation mention?
A: The README describes an NVIDIA GPU setup with about 12 GB of memory. It also mentions an MLX implementation with a native macOS app, without specifying its hardware requirements.
Q: What license applies to the weights?
A: The weights are released under the Apache License 2.0.
