Open-weight speech and local multimodal inference

Higurashi releases Haru-4B open weights for real-time, full-duplex speech

The Apache-2.0 base checkpoint listens and speaks simultaneously, with openly available local inference code and carefully qualified latency claims.

Open Model Weights published17 Nov 2026
Primary sourceHigurashi AI — Haru-4B technical announcement and official model card
Source published2026-11-13
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original dark emerald and electric blue vector artwork showing two synchronized neon audio waveforms moving through a glass dialogue nucleus.
Original conceptual depiction of simultaneous input and output audio, not measured Haru-4B spectrograms. AI-generated editorial illustration; not a documentary photograph.

Higurashi AI has released Haru-4B, a speech-to-speech foundation model built to listen while it talks. The November 13 technical announcement links both downloadable weights and inference code, making the work inspectable in a way that a hosted voice demonstration alone would not be. Haru is designed for interruptions, backchannels and overlapping conversation, where a rigid chain of transcription, language generation and speech synthesis can sound unnatural. Primary source ↗

The architecture combines three synchronized streams: the user's incoming audio, the model's outgoing audio and generated text. A Qwen3.5-4B-based language model processes those streams in continuous audio frames at 7.5 Hz, roughly one frame every 133 milliseconds. Its speech head generates a continuous acoustic representation with a single sampling step rather than a separate large acoustic transformer. This avoids some of the serial processing overhead of conventional voice pipelines, though the full deployed latency still depends on buffering and hardware.

Higurashi says its training involved more than 40 million hours of real speech spanning over 160 languages, with nine main target languages including English, German, French, Japanese and Chinese. On an Apple M1 Max machine, its local implementation reportedly achieves about 65 milliseconds of end-to-end system processing latency. Crucially, the developer excludes audio framing from that figure; framing adds up to another 133 milliseconds, and that result is not a universal measure of perceived conversation latency.

The model's Hugging Face card declares Apache-2.0 for the weights and links PyTorch, MLX and vLLM-Omni inference paths. It suggests around 12 GB of GPU memory for a supported NVIDIA configuration. But Haru is a pretrained base, not a reliably factual or instruction-tuned assistant. These distinctions matter for deployment: open weights enable research and adaptation, while evaluations of practical dialogue quality, safety and hardware behavior still require independent work.

OMW REGISTRY WATCH

What this changes for the evidence layer

Verify the published higurashi-ai/Haru-4B weight repository, actual file inventory and license revision before a canonical model intake. Classify it as a pretrained speech research base, not an instruction-tuned voice assistant. Do not report the vendor's Mac latency excluding audio framing as an independently measured response-time guarantee.

PRIMARY SOURCE

Higurashi AI — Haru-4B technical announcement and official model card

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗