Summary
According to a post on X by @HigurashiAI, Japanese startup higurashi has released Haru-4B, a pretrained base model that listens and speaks at the same time, under the Apache 2.0 license. The post says the company's blog describes the model as built on Qwen3.5-4B and pretrained on more than 40 million hours of real speech across 160+ languages, with no synthetic dialogues. It also reports about 65 ms end-to-end latency on a MacBook, with weights on Hugging Face and PyTorch and MLX code available. These are the company's own claims, and VQV has not independently checked them.
AI-assisted summary based on the listed source.
What happened
Japanese startup higurashi released Haru-4B, a pretrained base model that listens and speaks at the same time, under Apache 2.0. Its blog says it is built on Qwen3.5-4B, was pretrained on more than 40 million hours of real speech across 160+ languages with no synthetic dialogues, and runs at about 65 ms end-to-end...
Why it matters
Developers building voice assistants and real-time speech tools could test an openly licensed full-duplex base model as a starting point. The claimed 65 ms latency on a consumer laptop would matter most for teams that need responsive conversation on local hardware, if it holds up. Watch for independent benchmarks and community testing of the released weights to confirm the reported performance.
Signal Intelligence
Signal Strength 91%
Technical label WATCH
Public Interest 40
Category OPEN SOURCE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 44
Novelty Interest Score 94
Consequence Score 18
Curiosity Score 48
Shareability Score 53
Why this is here
VQV surfaced this signal because it is recent, relevant to AI Voice, connected to Scobleizer News Pulse (X).
