SAM 3 (Segment Anything Model 3)

What it is

SAM 3 is a unified model that detects, segments and tracks objects in both images and video. Meta announced it on 19 November 2025 together with SAM 3D, and it shared the SAM 3 model weights.^[1] The model accepts short open-vocabulary text prompts, such as *red baseball cap*. It also accepts image exemplars and visual prompts such as masks, boxes and points.^[3][1]

"Open-vocabulary" means the prompt is not limited to a fixed list of object names. The user can describe an object in ordinary words, and the model tries to find it.^[3]

Origins

The paper, *SAM 3: Segment Anything with Concepts*, had 38 authors. The first authors were Nicolas Carion and Laura Gustafson. It was submitted to arXiv on 20 November 2025.^[2]

Its training data came from a data engine that combined SAM 3, human annotators and AI annotators. The data covers over 4 million unique concepts.^[3][2] Meta released the model alongside the SA-Co (Segment Anything with Concepts) benchmark, which is a standard test set for this kind of task, and alongside SAM 3D, a pair of open models for reconstructing 3D objects and human bodies from a single image. Both can be tried in the Segment Anything Playground.^[3][1]

Main features

Promptable concept segmentation. Meta introduced this task: finding and segmenting every instance of a concept defined by a text prompt or an exemplar prompt. A single prompt such as a kind of object can therefore return many masks, one for each matching object in the scene.^[3]

Accuracy. Meta reported that SAM 3 doubles the accuracy of existing systems on promptable concept segmentation, in both images and video. Users preferred its outputs over the strongest baseline, OWLv2, at about three to one.^[2][3]

Speed. On an H200 GPU, the model processes an image in about 30 milliseconds, even with more than 100 detected objects.^[3]

Architecture. The text and image encoders come from Meta's Perception Encoder. A DETR-based detector finds the objects, which is a design that uses a Transformer to predict objects directly. A tracker built on the memory components of SAM 2 follows them through video.^[3]

Products. Meta named two uses of the model: effects in Instagram Edits, and Facebook Marketplace's "View in Room" feature.^[3]

Its place in development

SAM 3 builds directly on SAM 2. Its tracker reuses SAM 2's memory components, so the earlier work on following objects through video carried forward into this model.^[3]

Meta released SAM 3.1 on 27 March 2026. It added object multiplexing, which lets the model track up to 16 objects in one forward pass. On one H100 GPU, throughput for videos with a medium number of objects rose from 16 to 32 frames per second.^[3]

For a learner, the change from SAM 1 to SAM 3 is easiest to see in the prompts. The first model worked from points, boxes and masks. SAM 3 adds a way to ask for a whole category by name, and that moves segmentation closer to how people describe what they are looking for.^[3]