Haru-4B: Open 5.3B-Parameter Audio-to-Audio Model Released on HuggingFace

On November 12, 2026, the model higurashi-ai/Haru-4B with the audio-to-audio task was published on HuggingFace. It is a base model for continuous conversational dialogue: it listens and speaks at the same time, like a person on a phone call. Every 133 ms the model takes in a frame of user audio and outputs a frame of its own audio and text. The weights take up 9.8 GB in safetensors format, and the license is apache-2.0. The author describes the model as multilingual: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean are claimed.

What it means

Haru-4B is not text-to-speech but a full-fledged audio-to-audio dialogue, so it does not overlap in task with the tools from the audio generators section. The closest related tools solve a different problem. XTTS from Coqui is open-source text-to-speech, but the weights are under the Coqui Public Model License 1.0.0, which permits non-commercial use only. The largest weights file is 1.7 GB. Haru-4B is licensed under apache-2.0 for both code and weights, which is noticeably broader in its terms. YuE is a music generator, not a speech one. Its weights are also under apache-2.0, but the largest file is 4.6 GB, and 11.6 GB in total. The author specifies 24 GB of VRAM to run it. Haru-4B, as described by the author, requires about 12 GB on an NVIDIA GPU. Fish Speech is text-to-speech with weights under cc-by-nc-sa-4.0 and code under the Fish Audio Community License, which restricts commercial use to revenue of up to 500,000 dollars per year. The largest weights file is 1.2 GB. Haru-4B stands out by its task: it is not speech synthesis from text but a continuous dialogue model that decides for itself when to speak and when to listen. For a reader looking for a tool to voice text, it is no replacement for XTTS or Fish Speech. For live voice communication scenarios, it is a separate class of tool that the reference has not had until now.