Meta открыла SAM 3 и DINOv3 для Genesis Mission: анализ научных данных за 15 минут

Meta opened SAM3 and DINOv3 for Genesis Mission: scientific data analysis in 15 minutes

On July 21, 2026, Meta released two computer vision models into open access — Segment Anything Model 3 (SAM3) and DINOv3 — for the US scientific initiative Genesis Mission. Together they cut the analysis of scientific images from months of manual work down to about 15 minutes.

What Genesis Mission is

Genesis Mission is a White House initiative launched by the US Department of Energy (DOE) in late 2025 to accelerate scientific discovery with AI. Its flagship project, SYNAPS-I (SYnergistic Neutron And Photon Science – Intelligence), brings together five national laboratories and turns the analysis of scientific instrument data from a months-long bottleneck into near-real-time processing.

Key participants and figures of the project:

• Five DOE laboratories: Berkeley Lab, Argonne, Brookhaven, Oak Ridge, and SLAC
• About 60 researchers on the SYNAPS-I team
• Two open Meta models: SAM3 and DINOv3
• The fine-tuned pipeline runs on 300 NVIDIA A100 GPUs at the NERSC computing center
• Result: an annotated 3D volume in about 15 minutes

How SAM3 and DINOv3 work

The two models split the work: DINOv3 finds structures in an image, and SAM3 outlines them with pixel-level precision. DINOv3 is a self-supervised vision model, meaning it learns to recognize visual patterns without manual human labeling. SAM3 outlines object boundaries as accurately as a human scientific annotator, but does it in seconds instead of hours.

The combined pipeline was fine-tuned on scientific images and deployed on 300 A100 GPUs at NERSC. According to the Meta AI blog, the system produces fully reconstructed and semantically labeled three-dimensional volumes in about 15 minutes.

Where it's already being used

The first test case was analyzing grapevine drought resistance from micro-CT scans. The pipeline builds a 3D model of the vine stem and automatically locates xylem vessels — the tubes that carry water. Previously, such annotation took about a month per time step; now the plant's response to drought can be studied nearly in real time.

The scale of the data explains why this matters. Berkeley Lab operates the Advanced Light Source (ALS) synchrotron: after its upgrade, it produces tens of petabytes of data per year — equivalent to 2 million hours of HD video — and its detectors capture up to 100,000 images per second. That flow cannot be processed manually.

"Artificial intelligence, advanced computing and experimental systems

are now analyzing data as it's generated and steering experiments in real time, replacing slow manual steps with adaptive, automated decision-making," — Dario Gil, US Deputy Secretary of Energy.

Why model openness matters

The open license lets national laboratories download the models, fine-tune them, and deploy them inside secure government infrastructure — without relying on external clouds. For science, this is critical: it keeps unpublished research data under the control of all five SYNAPS-I laboratories until publication.

What it means

General-purpose AI models from industry are moving into fundamental science: what used to slow down discoveries by months is now counted in minutes. Open access to SAM3 and DINOv3 gives government labs a world-class tool they can run on their own servers.

Frequently asked questions

Which Meta models are used in Genesis Mission?

Two open computer vision models — Segment Anything Model 3 (SAM3) and DINOv3. DINOv3 finds structures in an image, and SAM3 outlines their boundaries with pixel-level precision.

Are SAM3 and DINOv3 free?

Yes, Meta released both models into open access. They can be downloaded, fine-tuned on your own data, and deployed inside secure infrastructure without external cloud services.

How much faster is the data analysis?

Annotating xylem vessels in grapevine images used to take about a month per time step. The new pipeline, running on 300 A100 GPUs at NERSC, produces a finished 3D volume in about 15 minutes.

*Meta is recognized as an extremist organization and is banned in Russia.