Segmentation models have always made you point at one thing. SAM 3 flips the unit of work: hand it a noun — "striped umbrella," "person wearing red" — and it returns every matching instance across an image or a whole video, tracking each through time. The hard part was never drawing masks; it was teaching a model what a concept means at the scale of millions of them.
Key Capabilities
- Open-vocabulary, exhaustive segmentation. A text phrase or visual exemplar yields all matching instances, not a single clicked object — turning an interactive tool into something closer to a dataset-labeling engine.
- One model for images and video. An 848M-parameter design shares a vision encoder between a detector and a decoupled tracker, with a "presence token" that disambiguates near-identical prompts and cuts false positives on confusable concepts.
- Built on scale. An automated data engine annotated 4M+ unique concepts; on the SA-CO benchmark of 270K concepts the model reaches roughly 75-80% of human accuracy — open-vocabulary recall solid enough for real auto-labeling.
Great Fit / Look Elsewhere
Great fit if you need to find and track every member of a category — every car, every cell, every logo — across frames, or to bootstrap segmentation datasets without clicking object by object. Look elsewhere if you need lightweight on-device inference (848M parameters is heavy) or pixel-perfect single-object editing, where SAM 2-style interactive point prompting stays leaner.
