Amazon Web Services has made OpenFold3-preview, an open-source biomolecular structure prediction model, deployable directly from Amazon SageMaker JumpStart, according to a November 13 post on the AWS Machine Learning Blog. The move puts one of computational biology's most consequential model families behind a managed cloud endpoint — and hands research teams an openly licensed alternative to proprietary structure-prediction systems.
OpenFold3 was developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University. Like other open-science efforts that keep the latest AI research accessible to everyone, the project publishes far more than a black-box API: the code and model weights ship under the Apache 2.0 license, and the developers have published the training data and methods as well.
What OpenFold3 Can Predict
OpenFold3 is an all-atom structure prediction model. According to AWS, it predicts the joint 3D structure of a set of molecules directly from their sequences — and it is not limited to single protein chains. A single prediction can fold multi-chain protein complexes, protein–nucleic acid assemblies, and protein–ligand complexes, which covers a large share of the molecular interactions that matter in drug development.
That capability matters because structure prediction has quietly become core infrastructure for computational drug discovery. AWS's post notes that teams now rely on it for target identification, antibody and protein engineering, and structure-based virtual screening — the early-stage work that determines which drug candidates ever reach a wet lab.
An Open License Is the Point
The Apache 2.0 licensing is what distinguishes OpenFold3 in a field crowded with closed frontier systems. Research organizations and pharmaceutical companies can inspect the model, fine-tune it on proprietary data, and build commercial products on top of it under the license's terms, AWS said. For an industry where molecular structures and screening data are closely guarded trade secrets, a model whose weights and training methods are published gives teams a level of auditability that closed APIs cannot match.
AWS describes OpenFold3-preview as a research preview, which is worth keeping in mind: this is an early, evolving deployment rather than a finished product. But even a preview distribution of an open, all-atom folding model lowers the barrier for academic labs that cannot negotiate enterprise licenses with proprietary model providers.
Why It Runs on Asynchronous Inference
The infrastructure story is as interesting as the model itself. Most JumpStart models — large language models and embedding models among them — deploy to real-time endpoints that must return a response within a single request capped at 60 seconds. Structure prediction does not fit that mold. Folding time grows with the number and length of the chains in a query, and real-world proteins and complexes often take several minutes to predict.
For that reason, OpenFold3 on JumpStart runs on SageMaker Asynchronous Inference. AWS highlights two properties of asynchronous endpoints that suit scientific workloads: they queue bursts of requests internally, so a lab can submit a large batch of sequences without building its own job queue, and they can scale to zero instances when the queue is empty, which avoids paying for idle GPU time between experiments. The model runs on G6 GPU instances, and AWS cautions that G6 capacity can be limited in some regions — deployments that fail with an InsufficientInstanceCapacity error should retry later or choose a different supported instance size.
There is also a cost warning baked into the announcement: SageMaker endpoints are billed per instance-hour while running, including while idle. AWS's post tells users to follow its cleanup steps or configure scale-to-zero to avoid unexpected charges — practical advice for research budgets that rarely tolerate surprise cloud bills.
What It Means for AI-Driven Science
The distribution deal is a signal about where open science and commercial cloud infrastructure are converging. OpenFold3's developers built a model designed to be inspected, fine-tuned and extended; AWS built the serving layer that handles the awkward operational realities — multi-minute inference, bursty batch workloads, idle-time costs — that stop most labs from self-hosting. Together, they make a capability that recently required deep ML engineering expertise available to a standard research team with a cloud account.
The competitive landscape is also worth watching. Structure prediction models sit at the center of a drug discovery arms race that includes some of the best-funded AI labs in the world, and open alternatives change the negotiating position of everyone in that market. If a pharmaceutical company can fine-tune an Apache 2.0 model on its own data instead of renting a closed system, the economics of AI-assisted drug discovery shift meaningfully.
None of that resolves the scientific caveats. Predicted structures are computational hypotheses, not experimentally determined facts, and AWS's own post frames the deployment as a research preview for exactly that reason. But the direction is clear: the open-weight ecosystem that reshaped natural language processing is now reaching the molecular sciences, and the cloud providers are racing to host it.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
