SAM 3

Meta's third Segment Anything Model moves from clicking on single objects to promptable concept segmentation: given a short text phrase or an exemplar box, SAM 3 detects, segments, and tracks every instance of that concept across an image or video. Under the hood, an 848 million parameter model pairs a DETR based detector with a transformer tracker over a shared vision encoder, and the interactive point and mask refinement of earlier SAM versions still works. Meta released the code, checkpoints on Hugging Face behind an access request, and the SA-Co benchmark covering 270,000 unique concepts; a SAM 3.1 update in March 2026 added shared-memory multi-object tracking. Running it locally requires Python 3.12, PyTorch 2.7 or newer, and an NVIDIA GPU with CUDA 12.6 support. The custom SAM License permits commercial use, modification, and redistribution, with restrictions on military and weapons applications. Since the November 2025 release the repository has drawn over 11,000 GitHub stars, and the model quickly became the default open vocabulary segmentation baseline.