Tier 4 Deeply open
What the publisher claims: open source, open weights
The one-liner: Open chess self-play at scale, with copyleft code and community-run training.
The six questions
Engine source, binaries and network weights are downloadable.
GPLv3 grants use, modification and redistribution but imposes copyleft on the code, and the weights ship under project terms rather than an explicit licence.
Builds with CUDA, ROCm, ONNX Runtime, Metal or plain CPU BLAS.
The self-play training pipeline and run history are documented in the repository and on the project site, but original training data is only partly published.
The community has reproduced strong networks from the published code and self-play, though exact nets are not guaranteed.
LeelaChessZero/lc0 shows 585 merged PRs, 223 from external contributors, with community-run self-play infrastructure.
License
GNU GPL v3 for the engine, with an additional GPL section 7 permission for CUDA-linked builds; network weights distributed freely by the project OSI-approved — license source ↗
- Copyleft obligations on redistributed engine code
What is actually released
- engine, self-play and training code
- trained network weights
- self-play training games
- training-run documentation
Why it matters
A decade-long, still-running open training run, which is the strongest evidence that decision systems can stay open in practice.
Sources
- https://github.com/LeelaChessZero/lc0/tree/master/src
- https://raw.githubusercontent.com/LeelaChessZero/lc0/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.