AWS Adds OpenFold3 to SageMaker JumpStart: Protein Structures Without Your Own GPU Cluster
Running a protein structure prediction model takes GPUs, a working CUDA and PyTorch stack, heavy weights, and a server that can handle requests lasting minutes. AWS has packaged the open OpenFold3 into SageMaker JumpStart, so you can deploy it in your own account with a few lines of code. Requests go to S3, are queued for up to an hour, and the endpoint scales down to zero when idle. For biotech teams, this is a way to compute structures without building their own infrastructure.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Predicting the spatial structure of molecules has become a working tool in drug discovery, used for target selection, antibody design, and virtual screening. But running such models yourself is inconvenient. You need graphics cards, a working CUDA and PyTorch stack, large weight files, and a server that doesn't fail when a single request takes minutes rather than milliseconds. AWS has offered a ready-made solution: the open OpenFold3-preview model is now available in Amazon SageMaker JumpStart.
What the model is and what it does
OpenFold3 was developed by the OpenFold consortium and the AlQuraishi lab at Columbia University. It is an all-atom model: it predicts the joint 3D structure of a set of molecules from their sequences. It is not limited to a single protein chain. In one pass, you can assemble a complex of several proteins, a protein bound to DNA or RNA, or a protein with a small molecule such as a drug. Such a molecule is specified as a SMILES string.
The output consists of structures in mmCIF format and confidence metrics: pLDDT, PAE, and PDE. By default, the model returns five candidates per request, and a summary file includes a sample_ranking_score used to pick the best one. For complexes, the key number is ipTM, which shows how reliably the chains and ligands are positioned relative to one another. pLDDT values above roughly 70 usually indicate a reliable local structure.
The code and weights are released under Apache 2.0, and the training data and methods are also published. Pharmaceutical companies and researchers can fine-tune the model and build their own products on it. A caveat: the authors themselves describe OpenFold3-preview as a research preview, so predictions should be double-checked for critical decisions.
Why asynchronous endpoints
Most models in JumpStart, such as language and embedding models, run on real-time endpoints, where the response must come back within a single request, with a limit of 60 seconds. For folding, that is not enough: computation time grows with the number and length of chains, and real complexes often take minutes to process.
That is why OpenFold3 runs through SageMaker Asynchronous Inference. The flow works like this: you place a JSON request in S3 and call InvokeEndpointAsync, which returns immediately. The endpoint queues the job and processes it, with timeouts of up to an hour. The result, or an error description, appears in the S3 location you specified, and you can learn it is ready by polling or through Amazon SNS. A malformed request, such as one with an incorrect chain, still produces a result object, only with an error status.
This design has two useful properties for science. A built-in queue absorbs bursts, so a large batch of sequences can be submitted without building your own job system. In addition, the endpoint can scale down to zero instances when there is no work, so you don't pay for idle time between experiments. This matters because SageMaker bills by instance-hour, including idle hours.
Hardware, parallelism, and privacy
Three instance types are supported: ml.g6.2xlarge, ml.g6.4xlarge (the default), and ml.g6.12xlarge. A single prediction uses one GPU, so a multi-GPU machine does not speed up an individual calculation. However, on ml.g6.12xlarge with four graphics cards, you can run four requests at the same time by setting max_concurrent_invocations_per_instance=4. According to AWS's measurements, for large proteins the cost of one prediction on ml.g6.12xlarge is close to that on ml.g6.4xlarge, although the outcome depends on the input data and region.
The practical choice is as follows: 2xlarge or 4xlarge is enough for individual experiments, 12xlarge suits a team with batches of requests, and for high-throughput screening it is combined with autoscaling. If G6 capacity in a region is insufficient, you can scale out across several 4xlarge instances.
On privacy, AWS offers several measures. The endpoint runs with network isolation, and requests and results are stored in S3, so for proprietary sequences the recommendation is a dedicated bucket encrypted with your own KMS key, a strict access policy, and a lifecycle rule to delete unneeded files. You can also supply your own multiple sequence alignments (MSAs): they improve accuracy, especially for complexes, and they do not leave your environment.
What this means for teams
For biotech startups and small research groups, the main point is that the entry barrier is lower. Instead of weeks spent setting up infrastructure, you need an AWS account, a quota for a G6 instance, and a sample notebook from the model card. There are limitations too: the model remains a preview, G6 capacity can be scarce, and some calculations require MSA preparation. Before launching, check your quotas, configure permissions according to the principle of least privilege, and remember to delete endpoints or enable scale-to-zero.
- Nov 21, 2026AG-UI 1.0 gives AI agents a common language for talking to apps: 31 event types
- Nov 21, 2026OpenAI launches ChatGPT Wallet: agents buy without ever seeing your card number
- Nov 21, 2026US Advertisers Spent $70.7 Million on ChatGPT Ads in October, Up 127% From April
- Nov 21, 2026ONNX Runtime vanishes from GitHub: Microsoft hides repositories after package hack
Key Questions Answered
What the model is and what it does?
OpenFold3 was developed by the OpenFold consortium and the AlQuraishi lab at Columbia University. It is an all-atom model: it predicts the joint 3D structure of a set of molecules from their sequences. It is not limited to a single protein chain. In one pass, you can assemble a complex of several proteins, a protein bound to DNA or RNA, or a protein with a small molecule such as a drug. Such a molecule is specified as a SMILES string.
Why asynchronous endpoints?
Most models in JumpStart, such as language and embedding models, run on real-time endpoints, where the response must come back within a single request, with a limit of 60 seconds. For folding, that is not enough: computation time grows with the number and length of chains, and real complexes often take minutes to process. That is why OpenFold3 runs through SageMaker Asynchronous Inference. The flow works like this: you place a JSON request in S3 and call InvokeEndpointAsync, which returns immediately.
What are the key details on Hardware, parallelism, and privacy?
Three instance types are supported: ml.g6.2xlarge, ml.g6.4xlarge (the default), and ml.g6.12xlarge. A single prediction uses one GPU, so a multi-GPU machine does not speed up an individual calculation. However, on ml.g6.12xlarge with four graphics cards, you can run four requests at the same time by setting max_concurrent_invocations_per_instance=4.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.