Tier 3 Practical open-weight
What the publisher claims: open-weight
The one-liner: Permissively licensed audio foundation weights from Moonshot.
The six questions
Weights are downloadable from the publisher's Hugging Face repository.
MIT weights with Apache-2.0 and MIT code — permissive, OSI-approved terms with no revenue or use threshold.
Runs locally on chosen hardware with published inference code.
A technical report is cited for the method; training data and recipe are not published.
Limited.
MoonshotAI/Kimi-Audio is a public repo that has merged a handful of outside PRs, but Moonshot AI alone chooses the audio recipe, checkpoints, and release timing, with no public roadmap or community decision-making.
License
MIT for the weights; Apache-2.0 and MIT for the code OSI-approved — license source ↗
- Standard MIT and Apache-2.0 obligations only
What is actually released
- weights (7B Instruct)
- inference code
- technical report
Why it matters
Noteworthy because the surrounding Kimi text families carry custom terms: the audio branch shows the same lab choosing ordinary permissive licensing.
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.