Accessibility Adjustments

Use these optional tools to adjust reading and display preferences. These tools cannot resolve every accessibility barrier. Please contact the website owner if you need assistance.

  • Text adjustments
  • Content scaling 100%
  • Font size 100%
  • Line height 100%
  • Letter spacing 100%
  • Colour adjustments
  • Orientation adjustments

Haru-4B opens full-duplex speech weights for local research

Listen to this article

Higurashi has released Haru-4B weights and inference code for developers researching conversations in which an AI can listen and speak at the same time. The official announcement is dated November 13, 2026. This is a pretrained speech research model, not a ready-made factual assistant or a new hosted subscription.

What full-duplex changes

Instead of waiting for an entire spoken turn to finish, Haru processes incoming speech alongside its own outgoing audio and text. Higurashi describes synchronized streams running at 7.5 Hz, equivalent to an audio frame roughly every 133 milliseconds. That design targets interruptions, overlapping speech and short acknowledgments during a conversation.

The practical distinction is interaction, not simply faster narration. A useful evaluation would check whether the model stops appropriately when interrupted, resumes coherently and avoids speaking over important information. Those are proposed tests, not results established by the release.

Weights are available, but plan the complete runtime

The official Hugging Face model card identifies the weights as Apache-2.0 licensed. It calls Haru a base model for research and warns that it is neither a reliable source of facts nor an instruction-following assistant. English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean are the nine named languages.

The public file listing includes a 10.5 GB model.safetensors checkpoint within a roughly 10.6 GB repository. Download size is not the same as required working memory. The card names a Qwen3.5-4B text backbone, while Hugging Face displays approximately 5B parameters for the complete model; the product name alone should not be used to estimate hardware needs.

Hugging Face currently lists no Inference Provider deployment for this checkpoint. Public weights therefore should not be confused with a hosted API that is ready to accept production traffic.

Choose the Mac or GPU path before installing

The official repository provides PyTorch with a web demo, MLX with a native Apple Silicon app, and a vLLM-Omni serving integration. Its NVIDIA quick start estimates about 12 GB of GPU memory. These are documented setup paths, not a guarantee that every machine with that capacity will perform identically.

The demo requires a supplied reference voice: one clean speaker recording between three seconds and three minutes. No voice clips ship with Haru. Use recordings you have permission to use. The repository also offers a sharing option that routes demo traffic through Gradio servers; that is different from keeping a session on the local machine.

The 65 ms claim excludes audio framing

Higurashi reports about 65 ms median latency on an M1 Max MacBook using int8 weights and wired audio input and output. Its breakdown excludes audio framing, which can add up to another 133 ms. This is a vendor measurement for that configuration, not a universal conversation-response time or a ByteForward benchmark.

Before choosing a speech stack, separate conversation behavior, factual reliability, startup costs and sustained device performance. For a different task, generating prepared speech rather than overlapping dialogue, our ZONOS3-Mini local runtime guide explains its separate setup and licensing constraints.

ByteForward reviewed the public announcement, model files and documentation. We have not installed Haru or tested its voices, conversation quality or latency.

Jordan Reid
Jordan Reid

Jordan Reid is focused on AI tools, agents, developer products, and the way technology changes everyday work. Jordan approaches a launch from the user’s side of the screen. What can it actually help someone finish? The voice is practical, conversational, and skeptical of products that turn a simple job into five new settings. Coverage follows coding assistants, creative software, browser agents, and the workflows around them, with attention to pricing, permissions, setup, and the human work that remains.

Leave a Reply

Your email address will not be published. Required fields are marked *

Gravatar profile