What happened
Higurashi’s November 13, 2026 announcement introduces Haru-4B as a full-duplex spoken-dialogue foundation model. The company says it aims to make interaction feel more like natural conversation by learning behaviors such as backchannels, interruptions and turn-taking from real speech. It releases the model as a pretrained base for further fine-tuning and research; weights are on Hugging Face, code on GitHub, and the announcement also points to demos.
Haru-4B is built on Qwen3.5-4B. Higurashi describes three synchronized streams: incoming user audio, generated model audio and model text. Their embeddings are combined and processed through a shared language model, allowing simultaneous listening and speaking without interleaving each stream token by token. The company says this reduces context length to one third compared with interleaving while keeping the text stream uninterrupted. Audio is represented as continuous 64-dimensional acoustic latents at 7.5 Hz, or about one token every 133 milliseconds. A lightweight MLP head generates audio in one sampling step, using MeanFlow distillation rather than iterative flow sampling or an additional acoustic Transformer.
The company reports pretraining on more than 40 million hours of speech spanning over 160 languages, with nine primary languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian and Korean. It says the training mixture excludes LLM-generated dialogue and TTS-synthesized speech. More than 13 million hours of separated full-duplex dialogue data were used in the second training stage, with samples of up to 60 minutes. Higurashi says it used FastDialogueSidon to separate and restore mixed conversational recordings into aligned streams, and Sidon to restore much of the speech data.
For its local latency result, Higurashi reports about 65 milliseconds end to end on a MacBook Pro with Apple M1 Max, using a native Swift application and custom Metal kernels. The measurement includes encoding, model inference, decoding and wired microphone/headphone I/O, but excludes audio framing; that framing can add up to 133 milliseconds. The listed p50 total is 65.2 milliseconds for int8 weights. On an NVIDIA B300, the company reports about 15 milliseconds for audio encoding and decoding plus one language-model step. Haru was pretrained on 256 NVIDIA B200 GPUs for approximately eight days. The announcement does not state a price for the open weights or code.
What the archive adds
New Runtime’s earlier coverage of Block’s Dessa described a full-duplex speech model trained only on human conversation audio (October 24, 2026 note). Haru-4B is a new release, adding a different approach: an openly released pretrained base built on Qwen3.5-4B, with disclosed multilingual training scale and latency measurements. Earlier coverage of speech recognition and text-to-speech models is adjacent, but does not describe the same task.
Trend signal
This release touches New Runtime’s broader generative-media infrastructure theme, but does not yet establish the pattern that speech is becoming a cheap, low-latency API with pinned versions for applications. Haru is an open base model and the announcement reports research and inference results, not a hosted API, production service levels or per-use pricing. This does not yet fit a pattern New Runtime tracks.
Why now
Higurashi says it is already pretraining larger full-duplex models, including models with more than 100 billion parameters, and is working toward real-time video input. Haru-4B is positioned as an open foundation for research while that work continues; the company also says it plans to publish a more detailed technical report and training code.
Connections to other events
The release sits alongside New Runtime’s earlier coverage of Block’s full-duplex model, which also focused on continuous spoken interaction, though the training approach described for Dessa differs. It also connects to the broader shift from turn-based voice interfaces toward systems that can respond while a user is still speaking. Haru’s announcement demonstrates a model architecture and reports benchmarks; it does not report tool calling, session-state management or deployment by an application provider.
Why it matters
For a team prototyping a voice agent, Haru offers a potentially useful alternative to chaining speech recognition, a text model and speech synthesis. A shared model with synchronized input, output and text streams could make interruptions and conversational timing part of the model’s learned behavior rather than glue code between services. Open weights and code also make it possible to inspect and adapt the system, although the release itself is a pretrained foundation rather than a complete product.
The practical question is not just whether the model can speak quickly on one device. Teams need to test recognition and response quality in their target languages, under noise and overlapping speech, and measure the full interaction loop including buffering and audio framing. The published MacBook figure excludes framing, which can add up to 133 milliseconds; that distinction matters when comparing it with end-to-end measurements from other systems. Builders should treat the reported numbers as a starting benchmark, then evaluate latency and conversational quality on their own hardware and scenarios.
Limits and uncertainty
The latency figures are Higurashi’s own measurements, and the announcement does not provide an independent replication or comparative benchmark against other models. It does not establish production reliability, quality across all 160-plus training languages, or how well Haru handles noisy recordings and overlapping speech in deployment. The reported 65-millisecond figure excludes audio framing, and the company notes that framing may add up to 133 milliseconds. The release is a pretrained base model; the announcement does not state a commercial license, hosted access, serving cost or availability guarantees. Higurashi also identifies separation errors and noise as limitations of its training data approach.
Skill to build now
Build a small evaluation set of real, consented conversations in the language and acoustic conditions your product will use. Include interruptions, overlapping speech, backchannels, silence and noisy microphone input. Run the model on your intended hardware and record both conversational quality and full end-to-end latency, including framing, encoding, decoding and I/O. That will show whether a low model-step time translates into a responsive experience for your users.
Reader takeaway
Haru-4B makes an open full-duplex speech foundation available for experimentation, with unusually specific disclosures about architecture, training scale and latency. Those disclosures are a useful starting point, not proof of a production-ready voice agent. Measure the complete interaction on your target hardware and languages before building around the headline latency.