Lawrence Berkeley National Laboratory has plugged two of Meta’s open-weight vision models directly into its X-ray and neutron science pipeline, according to a Meta blog post published July 21. The project, called SYNAPS-I, fine-tunes SAM 3, Meta’s third-generation Segment Anything Model, and DINOv3, its self-supervised vision model, to turn raw beamline scans into labeled 3D structures without a person tracing boundaries by hand.

The dependency here matters more than the demo. Genesis Mission, the White House initiative the Department of Energy launched in late 2025 to push AI into federal science infrastructure, counts SYNAPS-I among its flagship projects. A slice of publicly funded US materials and biology research is now standardizing its image analysis on models built and open-sourced by one private lab, not a government-built or academic-consortium alternative.

The data problem behind this is real regardless of whose framing you use. Berkeley’s Advanced Light Source, known as the ALS, and sister DOE facilities generate data at a scale Meta describes as tens of petabytes a year. Detector speeds have jumped roughly a hundred-thousand-fold, from one frame every six seconds to about 100,000 frames each second. Segmentation, the task of drawing precise boundaries around structures such as cell walls or mineral grains, is the bottleneck. Extracting usable structure from a single dataset has consumed weeks of expert annotation time.

SYNAPS-I is a multi-lab effort spanning Berkeley, Brookhaven, Oak Ridge, Argonne, and SLAC. It pairs SAM 3’s pixel-level boundary detection with DINOv3’s sense of what a structure is and where it sits within a sample. The team fine-tuned both models on beamline imaging data and deployed the pipeline on 300 A100 GPUs at NERSC, the national supercomputing center. Meta says the output is a fully labeled 3D volume handed back to a researcher standing at the instrument, with total turnaround around 15 minutes.

That 15-minute figure is Meta’s own number, not an independently verified benchmark. The blog post cites no peer-reviewed comparison of the fine-tuned pipeline’s segmentation accuracy against expert-labeled ground truth. Its only evidence for the speedup is Meta’s own comparison: work that reportedly took about a month per dataset is now finished in about fifteen minutes. Readers should treat that as a vendor’s characterization of its own deployment until an outside lab publishes an accuracy or throughput study.

The most concrete demonstration so far is agricultural. Micro-CT scans from the ALS feed the pipeline, which automatically identifies xylem vessels, the microscopic tubes that carry water through a grapevine stem, and tracks how they change as drought progresses. That kind of structural tracking could eventually speed work on drought-resilient crops, though Meta’s post stops at the pipeline demonstration and stops short of a published crop-science finding.

Meta’s case for why a national lab chose its models over a closed API is control. SAM 3 and DINOv3 ship as open weights, so SYNAPS-I can download, fine-tune, and run them entirely inside secure government computing environments, keeping unpublished research data off external servers. That is a genuine advantage over a closed API model. It is also how an open-weight release becomes durable lock-in. Once five national labs and 60 researchers have built tooling around a specific pair of checkpoints, switching to a different vision backbone means re-validating an entire segmentation pipeline.

Teams building imaging pipelines on SAM 3 or DINOv3 should wait for DOE to publish independent accuracy figures before citing the 15-minute claim as a validated benchmark, and should weigh the same lock-in risk before standardizing internal infrastructure on one lab’s vision stack.

Published by Meta on July 21, 2026.