Meta releases SAM 3 and SAM 3D

Background

Earlier SAM models worked from a prompt that pointed at one object, such as a click. SAM 3 also accepts a description in words. A user can therefore ask for every object that matches the description, not only the one being pointed at.^[1] Segmentation means outlining the exact pixels that belong to an object, which is more precise than drawing a box around it. Tracking means keeping the same object identified from one frame of a video to the next.

What happened

On 19 November 2025 Meta released SAM 3, which detects, segments and tracks objects in images and video. It accepts text prompts, such as "red baseball cap", as well as visual prompts.^[1] Meta released SAM 3D on the same day. It is two open models that reconstruct 3D objects and human bodies from a single image.^[1]

Key details

Concept segmentation. A paper describing SAM 3 reported that it doubles the accuracy of existing systems on promptable concept segmentation. This is the task of finding and outlining every instance of a concept named in a prompt.^[2]

Training data. The training data covers over 4 million unique concepts.^[2] A concept is a category of object or thing named in a prompt, such as a type of hat, so broad coverage helps the model respond to many different descriptions.

Weights and benchmark. Meta shared the SAM 3 weights, a new benchmark called SA-Co, and the Segment Anything Playground. The playground lets people try the models without technical expertise.^[1][3] Weights are the trained numbers that make up a model, and sharing them lets others run the model themselves.

Products. Meta said SAM 3 and SAM 3D power the "View in Room" feature on Facebook Marketplace.^[3]

Open release. Both SAM 3 and SAM 3D were released as open models, so researchers and developers could download and test them.^[1]

Impact

SAM 3 extended the segment-anything line from pointing at objects to finding objects described in words. The 3D models added reconstruction of objects and people from one photograph. Meta also used the same models in one of its own consumer features, which showed that the research could be used in a shopping product as well as in labs.^[1][3]