Tier 4 Deeply open
What the publisher claims: open-source
The one-liner: DeepMind's game-playing method survives without DeepMind's weights, in permissively licensed code.
The six questions
All code is downloadable and pip-installable.
Apache 2.0 and MIT.
JAX and PyTorch code runs on any accelerator, including laptops.
Algorithms are documented in the Nature and Science papers, but no trained checkpoints or official weights are published.
These reimplementations are the practical reproduction path for MuZero-style systems, and independent groups have trained equivalent agents.
Community reimplementations accept outside PRs, but there is no vendor governance to join.
License
Apache 2.0 for DeepMind mctx and pgx; MIT for alpha-zero-general and comparable projects OSI-approved — license source ↗
- No official DeepMind MuZero weights exist; these are independent implementations
What is actually released
- mctx search library (MuZero, Gumbel MuZero, AlphaZero)
- environment suites such as pgx
- from-scratch self-play implementations
- papers
Why it matters
When the lab withholds weights, the method can still be reproduced, so openness can come from the field rather than the publisher.
Sources
- https://github.com/suragnair/alpha-zero-general
- https://raw.githubusercontent.com/google-deepmind/mctx/main/README.md
- https://raw.githubusercontent.com/google-deepmind/mctx/master/LICENSE
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.