Tier 3 Practical open-weight
What the publisher claims: open-weight, open-source, Apache 2.0
The one-liner: An 82-million-parameter Apache-2.0 voice model that runs almost anywhere.
The six questions
Weights and voice packs are downloadable without gating from the publisher's repository.
Apache-2.0, an OSI-approved permissive license with no revenue or use threshold.
Small enough to run in browser or on CPU; one of the most portable models in the dataset.
The model card documents architecture, voice inventory, and provenance of the prerecorded voice material; training data and recipe are not published.
Limited.
hexgrad/kokoro is an open development line with 28 non-bot outside merges (voice packs, ONNX support, bug fixes) and a large community Discord, but a single maintainer (hexgrad) sets direction and publishes the checkpoints.
License
Apache-2.0 OSI-approved — license source ↗
- Standard Apache-2.0 obligations only
What is actually released
- weights (82M checkpoint)
- voice packs across eight languages
- inference code and model card
Why it matters
Shows that permissive weights plus small size can matter more than transparency for real-world openness: this one actually ships into products.
Sources
- https://huggingface.co/hexgrad/Kokoro-82M
- https://kokoroweb.app/voices
- https://www.eachlabs.ai/blog/kokoro-82m-small-tts-model-big-voice
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.