Higurashi drops Haru 4B full-duplex speech model

Haru 4B proves real-time conversational voice AI does not require heavy multi-stage cascades or proprietary closed APIs. By open-sourcing the architecture under Apache 2.0, Higurashi gives developers a practical foundation to deploy responsive voice agents on local hardware.

• –Native full-duplex streams: Synchronizes user audio, model audio, and text embeddings every 133 ms within a single LLM, slashing context length to one-third of interleaved approaches while handling interruptions naturally.
• –Single-step continuous generation: Replaces multi-step acoustic transformers and discrete RVQ codebooks with continuous latents and MeanFlow distillation for one-step inference.
• –Local inference capability: Achieves ~65 ms end-to-end latency on Apple Silicon and consumer GPUs, significantly lowering hardware barriers for real-time speech systems.
• –Pretrained base limitations: Released as a foundational base model without instruction tuning or tool calling, requiring downstream fine-tuning for interactive assistant tasks.