Skip to main content

higurashiAudio modelWeights available

Haru-4B

PopularityRank 12 of 24 audio models

34out of 100first value

usage data only, range 20 to 48

Summary

Haru-4B is a higurashi full-duplex speech model for continuous audio conversations. It processes and generates speech with accompanying text output. The research model supports nine languages, including German, English and Japanese.

Weights and inference code are publicly available. The weights use Apache 2.0. A reference recording can condition the voice, and audio processing uses 24 kHz.

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 54 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from November 15, 2026 to November 20, 2026

Haru-4BGemini 3.5 Transcribe Live 35GPT-Realtime-Whisper 34Granite Speech 5.0 470M TurboCTC 33GPT-4o Mini Transcribe 33

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Checkpoint
higurashi-ai/Haru-4B
Source: Checkpoint
Model class
Full-duplex speech model with continuous audio input and output
Source: Model class
Parameters
The released weights contain 5,270,000,897 BF16 parameters. “4B” refers to the Qwen3.5-4B text backbone.
Source: Parameters

Input and output

Input
Continuous speech audio input; a short reference recording conditions the voice.
Source: Input
Output
Continuous speech and accompanying text
Source: Output
Languages
English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean
Source: Languages
Text backbone configuration
The Qwen text configuration lists 262,144 positions. This is not a published end-to-end context limit for audio dialogue.
Source: Text backbone configuration

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Weights license
Apache License 2.0
Source: Weights license
Release
Released as a research base model on November 13, 2026.
Source: Release
Access
Public weights and inference code for local use; the reviewed model card lists no hosted inference provider.
Source: Access
Shutdown
None announced
no source

Further details

Audio frames
24 kHz, 3,200 samples per frame, and 7.5 frames per second (133 ms).
Source: Audio frames