Tier 3 Practical open-weight
What the publisher claims: open-source, open-weight
The one-liner: The most widely deployed open speech recognition stack, under a permissive license.
The six questions
Every published size is downloadable from the publisher's GitHub and Hugging Face repositories.
MIT-licensed code and weights, an OSI-approved permissive license with no commercial or use conditions.
Runs locally on CPU or GPU through the published code with minimal dependencies.
The paper documents the weakly supervised method and the scale of the training set at 680,000 hours; the corpus itself is not distributed.
Limited: the approach is well described but the training data cannot be reassembled.
openai/whisper has GitHub issues switched off entirely and OpenAI merges a trickle of outside PRs (about 124 merged PRs lifetime), with no public roadmap and no way for non-OpenAI contributors to steer Whisper's direction.
License
MIT License (code and weights; model cards also carry Apache-2.0 metadata) OSI-approved — license source ↗
- Standard MIT obligations only; no revenue threshold or use restriction
What is actually released
- weights (all published sizes)
- inference and training code
- model cards
- evaluation results
Why it matters
Demonstrates that permissive licensing plus a documented method can carry a model to universal deployment even without releasing the training data.
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.