A Guth Labs publication ↗
Back to News

Haru opens a local, talk-and-listen voice model, not a finished assistant

Haru opens a local, talk-and-listen voice model, not a finished assistant

November 13, 2026 | guthlabs.newsroom

Higurashi announced Haru-4B on November 13, releasing a voice model designed to listen while it speaks rather than wait for alternating turns. The release adds a downloadable research foundation to the growing set of local voice tools. It does not establish that natural-sounding conversation is reliable assistance. Higurashi announcement

What is available

The Apache-2.0 weights use a Qwen3.5-4B text backbone. The model card lists nine languages and explicitly describes the release as a pretrained base model, not an instruction-following assistant. It warns that the system is not a reliable factual source. The card's automatic size display rounds the complete checkpoint to 5B parameters: the “4B” product name should not be treated as an exact total-parameter or memory specification. Model card and weights

The public repository provides PyTorch, an Apple-silicon MLX implementation with a native macOS app, and a vLLM-Omni serving plugin. Its NVIDIA quick start calls for about 12 GB of memory. Voice reference clips must be supplied by the operator, who must have the right to use them. Inference repository

Read the latency footnote

Higurashi reports approximately 65 milliseconds on an M1 Max MacBook using int8 weights and wired audio. That measurement excludes audio framing, which can add up to 133 milliseconds. These are different components of delay, not competing headline numbers. The company also says it trained on more than 40 million hours of speech; that is a developer-reported training claim, not an independent audit. Technical explanation

PC Watch independently reported the November 13 announcement and open-weight release. Its report does not supply an independent replication of the latency or reliability claims. PC Watch

Guth News analysis

Analysis: This is worth evaluating as a conversational-interface component, separately from the agent that decides what actions to take. A fluent interruption or reassuring voice should not itself authorize a tool call. A useful trial would measure interruption recovery, factual accuracy and end-to-end delay on their own hardware before connecting the model to product workflows.

There is also a concrete deployment boundary: the repository's optional share mode creates a public link through Gradio's servers, and anyone with that link can use the demo. Local execution and publicly sharing the demo are different privacy choices. Repository sharing instructions

Verification note: The announcement, model card, repository and independent report were directly read. No software was installed or benchmark run. The event date above is the dated announcement, not a claim that the repository was first created that day. This is newly covered recent news, not a breaking-news claim.

Sources and citations

Revision history

Published .