Haru-4B

// Overview

Haru-4B is a full-duplex spoken dialogue foundation model that higurashi, a Tokyo-based speech AI startup, published on Hugging Face (higurashi-ai/Haru-4B) on 12 November 2026, with inference code on GitHub and a technical blog post dated 13 November 2026. Full-duplex means it listens and speaks at the same time, the way people do on a phone call, instead of waiting for the user's turn to end. A text character card sets who the model plays, and a reference clip of one speaker — 3 seconds to 3 minutes of clean speech — sets its voice. It is released as a pretrained base model for research and fine-tuning: the model card says it does not follow instructions the way an assistant does and is not a reliable source of facts.

Every 133 ms (7.5 Hz) the model takes one frame of the user's audio and produces one frame of its own audio and text. The three streams — incoming user audio, outgoing model audio and model text — are added element-wise into a single position of a Qwen3.5-4B language model, so listening and speaking never wait for each other. Audio is not a set of discrete codes: the causal VibeVoice acoustic tokenizer turns 24 kHz audio into one 64-dimensional continuous latent per 3,200 samples, and a small four-layer MLP head trained with flow matching and distilled with MeanFlow writes each latent in a single sampling step. Text is aligned to audio in chunks, which higurashi says helps mitigate catastrophic forgetting.

higurashi pretrained Haru-4B on more than 40 million hours of real speech in over 160 languages — LLM- and TTS-generated dialogue excluded — plus more than 13 million hours of separated two-speaker dialogue produced with its FastDialogueSidon tool, on 256 NVIDIA B200 GPUs for about 8 days with Megatron-LM. Nine languages are primary: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. higurashi reports about 65 ms end-to-end latency (p50 65.2 ± 1.5 ms, excluding up to 133 ms of audio framing) on an Apple M1 Max MacBook with int8 weights in its native Swift app, and about 15 ms per step on an NVIDIA B300 with vLLM-Omni. The repository ships a PyTorch implementation with a web demo, an MLX implementation with a native macOS app, and a vLLM-Omni serving plugin.

// Strengths

• Native full-duplex at 7.5 Hz: user audio, model audio and model text share one language-model position per 133 ms frame, so the model can listen while it speaks
• Continuous 64-dimensional audio latents generated in one MeanFlow step by a small MLP head — no discrete audio tokens and no separate audio transformer
• Character control from a text card and voice control from a 3-second to 3-minute reference clip
• Nine primary languages (English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, Korean), pretrained on speech in over 160
• About 65 ms end-to-end latency on an Apple M1 Max with int8 weights, as reported by higurashi
• Apache 2.0 weights with PyTorch, MLX (native macOS app) and vLLM-Omni implementations

// Best for

• Reach for it as a base to fine-tune a voice agent that can interrupt, back-channel and be interrupted instead of taking strict turns
• Reach for it to research full-duplex spoken dialogue with an openly licensed checkpoint and inference code
• Reach for it to prototype a local voice character on an Apple-silicon Mac or a single NVIDIA GPU with about 12 GB of memory
• Reach for it to serve full-duplex speech at low latency through the vLLM-Omni plugin

// How to access