A product of Braketlab Inc.Moment-driven Charge Density Joint Embedding Predictive Architecture

Charge density in milliseconds,
not CPU-hours.

MoC-JEPA is a pretrained framework with two backbones — M1 for crystals, O1 for organic molecules. Upload a structure, get per-node latents, then a charge-density grid you can drop into PySCF as an SCF initial guess.

quickstart.sh

1curl https://api.moc-jepa.ai/v1/density:direct \
2  -H "Authorization: Bearer $MOCJEPA_API_KEY" \
3  -F "model=moc-jepa-o1-den" \
4  -F "structure=@caffeine.xyz" \
5  -F "format=cube" \
6  -o caffeine_density.cube
7# -> caffeine_density.cube  (MoC-JEPA-O1-DEN-v1.0.pt)

Pretrained family

One architecture, two chemistries

Each backbone is pretrained independently. The shipping Density Prediction Model binds to that backbone's major line; later heads reuse the same naming slot.

Materials / crystals

MoC-JEPA-M1-v1.0.pt

Pretrained on crystal datasets in the Materials Project family. Built for solid-state chemistry and condensed-matter problems.

API id: moc-jepa-m1
Inputs: .cif / .xyz
Density head: MoC-JEPA-M1-DEN-v1.0.pt

Organic molecules

MoC-JEPA-O1-v1.0.pt

Pretrained on single-molecule organic datasets with more complex covalent topologies. Built for drug-discovery workflows.

API id: moc-jepa-o1
Inputs: .pdb / .xyz
Density head: MoC-JEPA-O1-DEN-v1.0.pt

Two pretrained backbones

M1 is trained on Materials Project-style crystals. O1 is trained on single-molecule organics for drug-discovery workflows.

Latents, then density

Encode per-node latents as .h5, then run the bound Density Prediction Model to a .cube or .chgcar grid.

Drop-in SCF acceleration

Initialize PySCF from the predicted density instead of a minao guess — fewer SCF cycles, same DFT answer.