SAM 3 is Meta AI's 2025 foundation model for promptable concept segmentation, which finds, segments, and tracks every instance of a concept in an image or video from a short text phrase or an image example. It unifies detection, segmentation, and tracking in one model.
Released by Meta AI on November 19, 2025, SAM 3 is the biggest leap in the Segment Anything line, and it is open source. Where SAM and SAM 2 segment one object per prompt, SAM 3 introduces promptable concept segmentation (PCS): you give it a concept, an open-vocabulary noun phrase like "yellow school bus" or an image example, and it detects, segments, and tracks every instance of that concept across an image or video.
It still accepts the classic visual prompts, points, boxes, and masks, and adds an automatic mode, but the headline change is moving from "segment this object" to "segment all of this kind of thing," in one unified model.
Key takeaways
SAM 3 does promptable concept segmentation, finding every instance of a concept, not just one object.
It is promptable by open-vocabulary text and image exemplars, as well as points, boxes, and masks.
It unifies detection, segmentation, and tracking across images and video in a single model.
What SAM 3 provides
What SAM 3 adds over SAM 2.
What SAM 3 adds over SAM 2.
Dimension
SAM 2
SAM 3
Prompt
Click, box, point
Text phrase, image exemplar, plus visual prompts
Scope
One object per prompt
Every instance of a concept
Tasks
Segmentation and tracking
Detection, segmentation, and tracking, unified
How it works
You prompt SAM 3 with a phrase or an example, and it returns masks for every matching instance, tracked through video. Under the hood, a detector and a tracker share a single vision encoder (roughly 848 million parameters), with stronger handling of small objects, clutter, and overlapping regions, and it ships with the large SA-Co concept benchmark. FiftyOne already supports SAM 3 through its integrations, so you can run concept prompts over a dataset, store the resulting detections and masks as labels, and review or correct them in the app. For a walkthrough, see the FiftyOne SAM 3 guide at https://voxel51.com/blog/how-to-use-segment-anything-model-sam3.
Why it matters
SAM 3 collapses the gap between "label this" and "label all of these," which is exactly the bottleneck in scaling annotation, because you describe a concept once and get every instance masked and tracked. By making segmentation open-vocabulary and concept-level, SAM 3 turns labeling into a prompt-engineering problem. The phrasing of the concept, "bus" versus "school bus" versus "yellow school bus," now drives recall and precision, so the skill shifts from drawing masks to specifying concepts precisely, and ambiguous concepts are where it still needs human review.
Frequently asked questions
When was SAM 3 released?
Meta AI released it on November 19, 2025, as an open-source model.
What is new in SAM 3 versus SAM 2?
Promptable concept segmentation. You can prompt with text or an example and get every instance of a concept, not just one object.
Does SAM 3 work on video?
Yes, it unifies detection, segmentation, and tracking across both images and video.
What is SAM 3D?
A companion model released with SAM 3 that reconstructs objects and human bodies in 3D from a single image.