Tier 3 Practical open-weight
What the publisher claims: open-weights, open-source code
The one-liner: Full-duplex speech dialogue with one of the most detailed training descriptions in the audio batch.
The six questions
Weights for Moshi, Moshiko, and the Mimi codec are downloadable from the publisher's Hugging Face repositories.
Code is permissively licensed, but the weights sit under CC-BY 4.0 with attribution obligations and a research-use framing rather than an OSI software licence.
Runs locally on consumer GPUs through the published Python and Rust implementations, with real-time full-duplex dialogue.
The model card names the training data mix (Stack Exchange, pes2o, several Wikipedia dumps, 87.5% filtered CommonCrawl) and the compute used (127 DGX nodes, 1016 H100s); the corpus is not redistributed.
Limited: unusually detailed description of method and infrastructure, but the data itself is unavailable.
Kyutai is a nonprofit research lab publishing kyutai-labs/moshi with a CONTRIBUTING guide and a real base of merged outside PRs (about 20 non-bot external merges, plus separate community finetune repos), yet the lab in Paris sets the research agenda and cuts releases.
License
CC-BY 4.0 for model weights; MIT for the Python code and Apache-2.0 for the Rust backend — composite terms with no single OSI-approved software licence over the release — license source ↗
- Weight licenses are content licences rather than OSI-approved software licences, and some Kyutai checkpoints carry non-commercial variants
- The model card states the system is intended for research use and advises against professional or advice-giving deployment
- Attribution is required for redistributed weights
What is actually released
- Moshi and Moshiko weights
- Mimi neural audio codec
- Python and Rust inference code
- model card with training-data description
Why it matters
Shows that the transparency axis can score well even when the weight licence is a content licence — openness here is a spectrum across A, P, D, T, and R rather than one switch.
Sources
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.