A Japanese spoken dialogue model released on February 25, 2026 by the Research and Development Center for Large Language Models at Japan’s National Institute of Informatics (NII). It is built on “Moshi,” the English full-duplex dialogue model (7B) developed by Kyutai in France, with additional training on Japanese conversational speech data; the model is named LLM-jp-Moshi-v1. Its defining trait is full-duplex operation: it can return backchannels such as “yes” or “I see” while the user is still speaking, and it handles turn-taking and pauses in a way closer to human conversation. It is published under the Apache License 2.0, including the model weights, so it can be used commercially. Rather than a web service, it is an open-source model you run on your own GPU — and that fact shapes everything about how it is used.
Key Features
- Full-duplex spoken dialogue: Instead of a turn-based structure that waits for the other party to finish, it listens and speaks at the same time. Backchannels, interruptions, and turn-taking occur naturally, bringing conversational tempo closer to human exchanges
- Direct speech-to-speech processing: It does not go through the three-stage pipeline of speech recognition, text generation, and speech synthesis; it works with speech tokens directly. As a result, response latency is low and the intonation and emotional nuance of speech are less likely to be lost
- Additional training on Japanese dialogue data: Pre-trained on the podcast dialogue dataset J-CHAT (about 69,000 hours) and fine-tuned on corpora including LLM-jp-Zoom1 (about 1,000 hours), a set of casual Zoom conversations collected by LLM-jp
- Commercially usable under Apache License 2.0: The model weights, the text tokenizer (Japanese SentencePiece), and the speech tokenizer (Mimi) are published on Hugging Face as
llm-jp/llm-jp-moshi-v1 - Improved quality over the earlier J-Moshi: According to NII’s announcement, the model scored higher than the existing Japanese Moshi-family model J-Moshi on speech intelligibility, naturalness, and semantic appropriateness
- Local execution: The official installation instructions assume Python 3.10 or later and a Linux GPU machine with 24 GB or more of VRAM. Running
python -m moshi.server --hf-repo llm-jp/llm-jp-moshi-v1starts the web interface
Pricing
| Plan | Price | Key features |
|---|---|---|
| Open source (Apache License 2.0) | Free | Model weights and tokenizers published. Commercial use permitted. Runs on your own GPU environment |
Pricing information is current as of August 2026. Please check the official site for the latest information. No hosted paid API is offered, so the cost of the GPU needed to run it is borne by the user.
Pros & Cons
✅ Pros
- Lets you keep full-duplex Japanese dialogue, including backchannels and turn-taking, on your own machine under a commercially usable license
- Direct speech-to-speech processing keeps response latency low and preserves the texture of how something is said
- Because the model runs in your own environment, audio data does not have to be sent to an external cloud
- Developed largely by universities and research institutes, with training data and evaluation results published in papers and presentation materials
⚠️ Cons
- It is not a web or smartphone app, so you need to prepare a GPU and set up a Python environment
- The official instructions assume a Linux GPU machine with 24 GB or more of VRAM, and macOS is not supported — this is not a setup you can try casually
- The developers themselves state that these are prototype models that may produce unnatural responses
- Because training centered on casual conversation data, it is not suited to task-specific business processing or use cases demanding factual accuracy
- Speech intelligibility and conversational continuity are listed as areas for future improvement, so product-grade quality is not guaranteed at this point
Comparison with Similar Services
| Comparison | LLM-jp-Moshi | J-Moshi | Moshi (Kyutai) | OpenAI Realtime API |
|---|---|---|---|---|
| Provider | National Institute of Informatics (NII) and others | Nagoya University Higashinaka Lab and others | Kyutai (France) | OpenAI |
| Main language | Japanese | Japanese | English | Multilingual |
| Delivery | Open source (local execution) | Open source (local execution) | Open source (local execution) | Cloud API (usage-based) |
| Full-duplex | Yes | Yes | Yes | Yes |
| License | Apache License 2.0 | Mainly for research use | Open source | Commercial API terms |
| Runtime environment | Linux GPU with 24 GB+ VRAM | GPU environment | GPU environment | None required (cloud) |
Who Is It For
- Researchers at universities and companies who want to study or prototype Japanese spoken dialogue systems
- Developers who want to run a conversational AI on-premises or in-house without sending audio data outside
- Engineers interested in the mechanics of human-like conversational tempo, including backchannels and interruptions
- Product developers looking for a baseline Japanese spoken dialogue model under a commercially usable license
- Students who want to learn academic evaluation methods for spoken dialogue and how training data is built
Summary
LLM-jp-Moshi is an open-source model that adapts the English full-duplex dialogue model Moshi to Japanese, and its value lies in making a natural conversational tempo — backchannels and turn-taking included — available under a commercially usable license. At the same time, running it requires a GPU environment with 24 GB or more of VRAM, and the developers position it as a prototype, so the realistic view is to treat it as a foundation for research and validation rather than something to drop straight into a product. It takes more effort than a hosted speech API, but the option to try full-duplex Japanese dialogue without letting data leave your environment is a valuable one.