Model compatibility & registry
Every model MoleculeDesk supports (or plans to support) is described by a manifest — a small, validated metadata file. The registry reads every manifest and exposes it to the CLI, so moldesk list always shows accurate, consistent information across very different upstream projects.
A manifest records things like:
- Category — e.g.
structure-prediction,sequence-design. - Status —
available,beta, orplanned(see below). - Runtime requirements — Python version, installer, package pins.
- Hardware — supported platforms (
darwin-arm64,linux-x64), and whether an NVIDIA GPU/CUDA is required or just recommended. - Input/output formats — what files the model expects and produces.
- License and source — upstream repository and a pinned, immutable revision.
What status means
available— install and run are both verified end-to-end across environments.beta— install and run both work today, but broader verification is still in progress before promotion toavailable.planned— listed for visibility, but installation/run support isn't complete yet.
A model being listed in the registry does not mean MoleculeDesk redistributes its weights — each model remains subject to its own upstream license and usage terms.
Checking compatibility
moldesk doctor compares your machine's Python, Docker, CUDA, GPU, RAM, and disk against what's declared in each manifest, so you know ahead of time whether a model can run on your hardware.
See also
- Boltz-2, DiffDock-L, BindCraft2, OpenDDE Preview, ProteinMPNN, LigandMPNN — current model pages.
- Contributing a new model adapter is documented in the repository's
docs/adding-a-model.md.