Haru-4B
A base model for live spoken conversation: it listens and speaks at the same time, like people on the phone, instead of taking turns. Every 133 milliseconds it takes a piece of the other person's audio and produces a piece of its own audio and text, so the latency is very small. English, Spanish, Portuguese, German, French, Chinese, Japanese and Russian are stated. It is released as a pretrained base for research: it holds a conversation, but it is not a ready assistant, so expect to tune it for your task. For the lowest latency the authors advise wired headphones and a microphone, as Bluetooth adds about 100 milliseconds.
- Developer
- Higurashi AI, not stated
- First release
- Nov 2026
- Latest release
- Nov 2026
- Sizes
- 4B (about 5.3B with the audio components)
- License
- Commercial use allowedApache 2.0
- Russian
- Supported
- Running
- On your own serverNeeds a GPU
- Industries
- Customer support, Media and production
What it does
- research and prototypes of voice assistants without pauses and turn-taking
- a base for fine-tuning a voice agent on your topic
- comparison with the usual scheme of recognition, text and voicing
Where it is used
Hardware requirements
Versions
- Haru-4B (предобученная база)
How to run it
I can set this up end to end: pick the model size, deploy it on your server and connect it to your systems.
Frequently asked questions
Can Haru-4B be used in a commercial project?
Yes. License: Apache 2.0. It allows commercial use, but it is still worth having a lawyer review the license before launch.
What hardware does Haru-4B need?
At minimum: One GPU with 16–80 GB — mid-size versions. Without a GPU the model is not practical. You can calculate the exact VRAM for your model size and context in the hardware calculator.
Does Haru-4B support Russian?
Yes, Russian is listed on the model card.
Where can I download Haru-4B and what does it cost?
The Haru-4B weights are open and free to download. You only pay for the hardware it runs on and for the setup. Source links are at the bottom of this page.
How I deploy it for clients
- SelectionI pick the model size for your task and hardware and test it on your examples.
- DeploymentI deploy it on your server or in a closed network and provide an API.
- Fine-tuningI fine-tune it on your data (LoRA) or connect a knowledge base — whichever is cheaper for the task.
- IntegrationI connect it to your CRM, ERP, bot, website or team chat and set up monitoring.
Similar models
A voice assistant that listens and speaks at the same time, with no delay for recognition and synthesis. Hibiki does simultaneous speech-to-speech translation between several European languages.
DetailsSpeech to textVoxtralMistral AI · FranceCommercial use with conditionsMistral's speech models: they understand audio, transcribe and answer questions about a recording. The Realtime version recognizes speech live and supports Russian; speech synthesis and a separate version for Arabic dialects are also available.
DetailsSpeech to textVibeVoiceMicrosoft · USACommercial use allowedMicrosoft speech models: long multi-voice dialogue synthesis, fast synthesis for live conversation, and recognition of long recordings split by speaker, including in Russian.
DetailsSource: huggingface.co/higurashi-ai/Haru-4B. Data checked against the model card on 22 Sep 2026. Have a lawyer review the license before commercial launch.


