Tier 5 Fully open / reproducible
What the publisher claims: open source, open weights, open training data
The one-liner: The closest thing in the batch to fully open and reproducible: code, data, weights and method.
The six questions
Code, prebuilt binaries, weights and training data are all downloadable.
Permissive licence with no field-of-use restrictions on the weights or the data.
Six backends (OpenCL, CUDA, TensorRT, ROCm, ONNX, CPU) plus prebuilt Windows and Linux binaries.
Self-play data, network architectures and method notes are published in the repository, and the public run keeps publishing checkpoints.
A skilled group can retrain from scratch; the project documents that one consumer GPU reaches strong amateur dan and, with months of tuning, superhuman strength.
Genuinely community-governed: lightvector/KataGo shows 153 merged PRs of which 118 are from non-affiliated contributors.
License
MIT-style licence (custom text with an additional disclaimer); network weights published by the project for free use OSI-approved — license source ↗
- Bundled third-party libraries under their own licences
What is actually released
- engine and self-play training code
- public training networks
- self-play game data (katagotraining.org and katagoarchive.org)
- method and network-architecture documentation
Why it matters
The cleanest positive case, a frontier-adjacent game-playing system where nothing material is withheld.
Sources
- https://katagotraining.org/
- https://raw.githubusercontent.com/lightvector/KataGo/master/LICENSE
- https://raw.githubusercontent.com/lightvector/KataGo/master/README.md
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.