Model Overview

Description:

AvgFlow as released in this repository is a diffusion transformer model with pairwise biased attention trained on 3-dimensional conformer data of small molecules. It can be used as a highly efficient molecular conformer generator.

This model is for research and development only.

License/Terms of Use:

GOVERNING TERMS: Use of this model is governed by the NVIDIA Open Model License. AvgFlow source code is licensed under Apache 2.0.

Deployment Geography:

Global

Use Case:

AvgFlow is a model for efficient molecular conformer generation. The model can be used in the pharmaceutical and chemical industries and acadamic research to generate the 3D conformers of small organic molecules for downstream tasks such as property prediction and virtual screening.

Release Date:

Github: 10/15/2025 via https://github.com/NVIDIA/avgflow

NGC: 10/15/2025 via

Reference(s):

Research paper: "Efficient Molecular Conformer Generation with
SO(3)-Averaged Flow Matching and Reflow", https://arxiv.org/abs/2507.09785

Model Architecture:

Architecture Type: Transformer with pair-biased attention

Network Architecture: Diffusion Transformer architecture with pair-biased attention, adaptive layernorm, and gating.

Number of model parameters: Two variants: 52 million and 64 million

Input:

Input Type(s): Text and 3D atom coordinates

Input Format(s): SMILES string and Cartesian coordinates

Input Parameters: 1D (SMILES string) and 3D (coordinates)

Other Properties Related to Input: Trained with maximum number of token (atom) as 200.

Output:

Output Type(s): 3D atom coordinates

Output Format: Cartesian coordinates

Output Parameters: 3D

Other Properties Related to Output: None Applicable

Software Integration:

Runtime Engine(s):

• JAX

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Ada

[Preferred/Supported] Operating System(s):

• Linux

Model Version(s):

AvgFlow v1.0

Training, Testing, and Evaluation Datasets:

The released model is trained on the public GEOM-Drugs (https://github.com/learningmatter-mit/geom)

Training Dataset:

GEOM-Drugs
Link: https://github.com/learningmatter-mit/geom

Data Modality:

• Molecule SMILES string
• Molecule 3D coordinates

Non-Audio, Image, Text Training Data Size (If Applicable):

• Approximately 300k molecules.

Data Collection Method by dataset:

• Semi-empirical quamtum mechanics calculation

Labeling Method by dataset:

• Automated

Dataset License(s): MIT License

Evaluation Dataset:

GEOM-Drugs

Link: https://github.com/learningmatter-mit/geom

Data Modality:

• Molecule SMILES string
• Molecule 3D coordinates

Non-Audio, Image, Text Training Data Size (If Applicable):

• Approximately 300k molecules.

Data Collection Method by dataset:

• Semi-empirical quamtum mechanics calculation

Labeling Method by dataset:

• Automated

Dataset License(s): MIT License

Inference:

Engine: JAX

Test Hardware Ampere (NVIDIA A100) / Ada (Nvidia A5880/A6000)

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards None.

Please report model quality, risk, security vulnerabilities or concerns https://developer.nvidia.com/support.