SageMaker JumpStart has added OpenFold3-preview as a ready‑to‑use asynchronous inference endpoint, delivering an all‑atom biomolecular structure model in a managed container that runs on GPU‑enabled instances. Engineers can now launch the model with a few API calls, avoid manual CUDA/PyTorch setup, and benefit from built‑in request queuing, autoscaling, and zero‑instance scaling.
OpenFold3 asynchronous inference
The model is identified by the JumpStart ID huggingface-proteinfolding-openfold3 and runs on the ml.g6.2xlarge, ml.g6.4xlarge (default), and ml.g6.12xlarge instance families. It operates behind SageMaker’s asynchronous inference workflow: a client uploads an input payload to Amazon S3, calls InvokeEndpointAsync, and receives an immediate token. The service then processes the job for up to one hour, writing results or error details back to a designated S3 location and optionally notifying via SNS.
Implementation considerations
Deploying the endpoint requires selecting a supported GPU instance type and configuring an S3 bucket for input and output. Because the endpoint is asynchronous, callers must handle polling or SNS notifications to retrieve results. The model accepts protein, DNA, RNA sequences, and small‑molecule SMILES strings; users can also provide custom multiple sequence alignments (MSAs) to improve accuracy while keeping raw sequences inside their account.
Operational impact
Running OpenFold3 on an asynchronous endpoint shifts the scaling model from always‑on to demand‑driven. SageMaker can automatically scale the number of GPU instances based on queue depth, and it can scale the endpoint to zero when the queue is empty, reducing idle costs. Multi‑GPU instances enable parallel processing of several prediction jobs, which is useful for high‑throughput screening pipelines that batch large numbers of sequences.
Security and data isolation
The endpoint is deployed with network isolation, meaning that inbound and outbound traffic is confined to the VPC and the model cannot reach the public internet. This isolation helps keep proprietary sequence data within the AWS account. Since the model and its checkpoints are released under the Apache 2.0 license, organizations can inspect the code, verify compliance, and, if needed, fine‑tune the model under the same license terms.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Practitioners can replace custom GPU clusters and ad‑hoc job queues with a managed SageMaker endpoint, gaining predictable cost control through autoscaling and zero‑instance idle periods. Security teams should verify that the VPC configuration aligns with internal isolation policies, and operations teams should monitor S3 input/output locations and SNS notifications for job health. Evaluating the trade‑off between real‑time and asynchronous endpoints will guide workload placement, especially for workloads that exceed the 60‑second real‑time limit.
