Amazon SageMaker JumpStart now offers a managed deployment path for the OpenFold3 research preview. The open-source model predicts joint three-dimensional structures for proteins, DNA, RNA, small-molecule ligands and complexes, and runs inside a customer’s AWS account.
The deployment uses SageMaker Asynchronous Inference because structure predictions can take several minutes, exceeding the 60-second limit of a real-time endpoint. A user uploads a request to Amazon S3, receives an immediate job reference and later retrieves the predicted structures and confidence scores from S3. The endpoint can queue bursts, run predictions across multiple GPUs and scale to zero when idle.
OpenFold3’s code, weights, training methods and data are available under Apache 2.0. Research groups can inspect or adapt the system for target identification, protein engineering, protein-ligand hypotheses and batch screening. Users can also supply their own multiple-sequence alignments, which may improve accuracy while keeping sequence-processing steps under their control.
The preview does not remove infrastructure or scientific limits. Teams still need suitable GPU quotas, storage permissions and careful protection for proprietary sequences and results. Endpoints incur hourly charges while running, and scale-to-zero must be configured to avoid idle cost. Predicted structures and confidence values remain computational hypotheses that require domain review and, where necessary, experimental validation.