How they use Nebius AI Cloud

Quantori trained a molecular-generation model on a Nebius cloud cluster with eight NVIDIA H200 GPUs. The research used 1.6 million ChEMBL compounds to teach the model to generate molecules with specified 3D shapes. Training ran for 1,500 epochs, approximately 13.5 days.

The pipeline combines an Equivariant Diffusion Model with Quantori’s pre-trained Structure Seer model. Shape descriptors guide molecular generation, and Structure Seer predicts bonds from the generated atom coordinates. Quantori reported valid 3D conformers with shapes similar to reference molecules. Shape similarity was evaluated using the shape Tanimoto Score, which measures molecular volume overlap. In a preliminary evaluation of 10,000 generated molecules, Quantori reported that more than 99.8% were unique relative to the training dataset, with an average shape similarity to the reference of 0.54. Its company account credits quick provisioning and easy configuration on Nebius with simplifying training and evaluation. In Nebius’s Q1 2025 earnings call, Daniel Bounds said the framework increased the number of molecules Quantori could model and enabled faster, scalable R&D for drug and materials discovery.

Quantori’s technical presentation walks through the training run and demonstrates generation on one Nebius GPU. Its company account reports generating 50 samples in approximately 90–120 seconds without parallelization or additional optimization. These results concern its molecular-design research project. Nebius also describes Quantori using containerized, GPU-accelerated OpenMM batch jobs for large-scale molecular simulations. Jobs are submitted from Python with automatic GPU provisioning. The webinar announcement plans a live protein-ligand simulation, a parallel parameter sweep and automatic delivery of trajectory and log outputs to object storage. Quantori’s announcement also plans CPU-heavy docking with automatic CPU provisioning from Python and real-time docking-assisted molecule generation. A joint webinar recap published by Quantori on April 24, 2026 also describes a demonstrated workflow from pocket detection to spatially aware generative ligand design, with serverless docking and no cluster setup or Kubernetes. Quantori’s joint webinar post names MLConfGen alongside serverless docking as enabling rapid iteration.