Tier 3 Practical open-weight
What the publisher claims: open source
The one-liner: Open method, closed model: MIT code with no weights to open.
The six questions
Full source, configs, environment wrappers and score logs are downloadable — but no trained weights exist from anyone, so what you access is the method, not a model.
MIT on the code.
Trains locally from configs on user hardware.
Nature paper and repo document the algorithm and hyperparameters in full; the DeepMind-scale training runs are not reproducible from the reimplementation.
A skilled group can rebuild a substantially equivalent system at reduced scale — reproducibility here means the method, not the artifact.
Danijar Hafner's MIT reimplementation is a personal open project with an active issue tracker, merged community pull requests, and downstream community forks that shaped the JAX version, though the roadmap rests with a single maintainer.
License
MIT OSI-approved — license source ↗
- The public repo is the author's reimplementation, explicitly unrelated to Google or DeepMind
- No pretrained weights released by any party
What is actually released
- training code
- configs.yaml per environment
- Docker setup
- environment wrappers
- score logs
Why it matters
Stretches the rubric honestly: Access measures artifacts that exist, and for decision models the artifact is often the method itself — a different shape of openness than a weight release.
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.