Meta's SAM 3 and DINOv3 power DOE Genesis Mission pipeline

A quietly interesting item from Meta's AI research blog: the company's open-weight SAM 3 and DINOv3 vision models have been fine-tuned by staff at four DOE national labs and deployed on 300 A100 GPUs to serve the opening phase of the White House's Genesis Mission, launched in late 2025. The pipeline they power, SYNAPS-I, stands for SYnergistic Neutron And Photon Science with Intelligence, and it runs across Argonne, Brookhaven, Oak Ridge and Lawrence Berkeley National Laboratory, with 60 researchers involved.

The concrete result Meta leads with is a grapevine study. Using micro-CT scans from the Advanced Light Source, the pipeline automatically identifies xylem vessels, the water-transport structures inside the stems, and tracks how they change as drought progresses. That kind of segmentation used to consume roughly a month of expert effort per time step. According to Meta, on the SAM 3 and DINOv3 stack it now takes about 15 minutes for a fully reconstructed 3D volume. That is the sort of speed-up that changes what an experiment is even for, moving analysis out of the write-a-paper timescale and into the tune-the-next-shot timescale.

Why this matters beyond the grapevines is the deployment pattern. DOE's light and neutron sources reportedly generate tens of petabytes a year, with Advanced Light Source detectors capturing 100,000 images per second, and the segmentation bottleneck has been human. Putting open-weight foundation models on lab-owned GPUs, rather than calling a hosted API, is the only version of this that works inside secure federal compute. Under Secretary Dario Gil framed it as replacing slow manual steps with adaptive, automated decision-making that can guide experiments in real time.

The honest caveat is that the reporting here comes from Meta itself and is thin on important detail. There is nothing on how the fine-tuning was done, who owns the resulting checkpoints, how results are made reproducible outside DOE compute, or why Meta's models were chosen over other open-weight vision options. Treat the 15-minute number as a headline claim about a specific grapevine workflow, not a general benchmark. The part worth watching is whether the same stack scales to the other Genesis Mission workstreams the national labs are queuing up behind it, because if it does, an open-weight vision model just became load-bearing infrastructure for U.S. science.