OpenFold3 Parameters¶
The list of parameters available with OpenFold3 are as follows
Checkpoint Name |
File Name |
Version compatibility |
Number of training steps |
|---|---|---|---|
openbind-2025-06-30-174k (default) |
|
|
174,000 |
openfold3-p2-155k (deprecated) |
|
|
155,000 |
openfold3-p1 (deprecated) |
|
|
78,000 |
How do I use an older set of parameters?¶
To use a deprecated set of parameters, install an older version of OpenFold3.
For example, to use OpenFold3p2 weights, you can use pypi to install a compatible version
pip install openfold3<0.5
Rerun setup_openfold after changing versions to ensure that the package has access to compatbile weights.
Which parameters is OpenFold3 using?¶
If a specific parmater file path is provided, OpenFold3 will use that file. Otherwise, OpenFold3 will first check default directories for previously downloaded parameters.
The following schematic describes the resolution of paths through searching default directories.
flowchart TD
START([Start]) --> A{"<i>inference-ckpt-path<i> provided?"}
A -- YES --> Z(["Use provided path: <i>inference-ckpt-path<i>"])
A -- NO --> B{"<i>inference-ckpt-name<i> provided?"}
B -- YES --> C["Use <i>inference-ckpt-name<i>"]
B -- NO --> D[Use default model: <i>openfold3-p2-155k<i>]
C --> E[Parameter path resolution]
D --> E
E --> F{"<i>$OPENFOLD_CACHE<i> set or <i>cache-path<i> specified"?}
F -- YES --> G{<i>ckpt_root<i> file exists in <i>$OPENFOLD_CACHE?<i>}
F -- NO --> H([Download parameters to <i>~/.openfold3<i>])
G -- YES --> I([Download parameters to $OPENFOLD_CACHE/ckpt_root path])
G -- NO --> J(["Download parameters to <i>$OPENFOLD_CACHE<i>"])
classDef decision fill:#dbeafe,stroke:#3b82f6,color:#1e3a5f
classDef terminal fill:#dcfce7,stroke:#22c55e,color:#14532d
classDef download fill:#fef9c3,stroke:#eab308,color:#713f12
classDef process fill:#f3e8ff,stroke:#a855f7,color:#3b0764
class A,B,F,G decision
class Z,I terminal
class H,J download
class C,D,E process