LLM-jp Moshi v1
LLM-jp Moshi v1 adapts Kyutai's 7B full-duplex Moshi model with Japanese dialogue audio. It supports spoken interaction rather than text-only completion.
Research release
Text and speech tokenizers have different origins
The release includes rinna's Japanese GPT-2 Medium SentencePiece text tokenizer and Kyutai's Mimi audio tokenizer. Its prototype notice warns that responses can be unnatural.LLM-jp Moshi v1 official model card
Test timing as part of conversation
Practical guidance. Measure interruption handling, response timing and audio feedback as well as transcript quality. These affect a full-duplex conversation in ways that a text completion benchmark cannot capture.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | llm-jp-moshi-v1 |
| Ancestry | kyutai/moshiko-pytorch-bf16; continued training for Japanese dialogue.LLM-jp Moshi v1 official model card |
| Training audio | J-CHAT, about 69,000 hours, and LLM-jp-Zoom1, about 1,000 hours.LLM-jp Moshi v1 official model card |
| License | Apache 2.0.LLM-jp Moshi v1 official model card |
| Published runtime requirement | Linux GPU machine with at least 24 GB VRAM; the card says macOS is unsupported.LLM-jp Moshi v1 official model card |
Related models
- KABURI-TTS
LLM-jp / NII LLMC · KABURI-TTS
Japanese two-speaker dialogue synthesis with separate audio channels, overlaps and backchannels.
Launch date not confirmedResearch releaseDownloadable - Llama-Mimi 8B
LLM-jp / NII LLMC · Llama-Mimi
An English speech-continuation model derived from Meta Llama 3.1 8B and Mimi audio tokenization.
Launch date not confirmedResearch releaseDownloadable