Haru-4B

by Higurashi · Model · Voice & audio

weak

Interesting open voice model, but the only hands-on test found it unreliable

Haru-4B is a full-duplex speech-to-speech dialogue model built on Qwen 3.5, open weights, Apache 2.0, claims nine languages

AccessOpen weightsNeeds~10.5 GBDownloadabout 10.5 GBLicenseApache 2.0

Why we say so

Fahd Mirza ran it locally on about 12 GB of VRAM and found it worked only in English and Japanese, not the nine languages Higurashi claims. Its demo answers were irrelevant, and with a custom persona it repeated itself and ignored questions. Higurashi itself calls it a base model for research, and its 0.32 s latency claim is untested by others.

Evidence

  1. Nov 19
    Architecture: audio sliced into ~133 ms steps, user voice, own voice and words fed to one LM predicting next word and voice slice together, continuous voice representation
  2. Nov 15
    Open full-duplex speech-to-speech voice model; generates speech directly without text step; about 12 GB VRAM; Apple Silicon support

Cost to run