Technology

SAM3

Meta’s unified foundation model for Promptable Concept Segmentation (PCS) that detects, masks, and tracks every instance of a visual concept across images and videos.

SAM 3 unifies detection, segmentation, and tracking into a single 848M-parameter architecture for Promptable Concept Segmentation (PCS). This model moves beyond the single-object clicks of its predecessors to handle open-vocabulary tasks: you can now isolate every instance of a concept (like "yellow school bus" or "person with backpack") across images and video frames simultaneously. It cuts through the complexity of multi-model pipelines by integrating a DETR-style detector with SAM 2’s temporal memory. For teams working in spatial AI, the companion SAM 3D model provides a direct path from single RGB images to high-fidelity 3D reconstructions.