higurashi Haru-4B Achieves 65ms Latency with 40 Million Hours of Voice Training
·2026.11.16 16:00
Key point
higurashi released Haru-4B, a full-duplex voice model trained on 40 million hours of real voice data.
Details
- Released Haru-4B, a full-duplex voice conversation model based on Qwen3.5-4B, as open source under the Apache 2.0 license
- Used over 40 million hours of real voice data in more than 160 languages, excluding synthetic data, as training data
- Recorded an end-to-end latency of approximately 65ms in a MacBook M1 Max (int8) environment and an inference latency of approximately 15ms on an NVIDIA B300 server
- Applied a Full Attention structure to one out of every four layers among the 32 layers, with the remaining layers using a Gated DeltaNet (Linear Attention) structure
- Utilized the VibeVoice causal acoustic tokenizer, converting 3,200 samples of 24kHz audio into a single 64-dimensional continuous latent vector
- Optimized the flow matching head for single-step generation via MeanFlow distillation, with a default guidance scale of 1.5
- Supports PyTorch, MLX, and vLLM-Omni plugins; checkpoints are approximately 10GB in size based on BF16
- Currently a voice-only model, with development in progress for full-duplex models with over 100B parameters and omni-modal models
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.