Meta Releases Concept-Based Segmentation Model SAM 3

On August 4, 2026, Meta AI Research released SAM 3 (Segment Anything with Concepts), a new model that segments and tracks every instance matching a specified “concept” in a single pass. Where earlier SAM models segmented one object at a time from a point or box prompt, SAM 3 accepts a text prompt or image example defining a concept, then finds, segments, and tracks every matching instance across an entire image or video.

What Changed

According to Meta’s official announcement, SAM 3 introduces a new task called Promptable Concept Segmentation (PCS). Given a text prompt (e.g., “person wearing a red baseball cap”) or an image exemplar, it detects, segments, and tracks every instance matching that concept across an image or video. The key difference from earlier SAM/SAM 2 versions — which focused on segmenting a single instance from a point-or-box prompt — is the ability to handle multiple objects at once, at the level of a concept rather than a single click.

The example below applies the concept “person wearing a red baseball cap” to the same photo to show how SAM 2 and SAM 3 behave differently. SAM 2 segments only the single instance tied to a specific point, while SAM 3 finds and segments every instance matching the concept across the whole image. There are two people wearing a red baseball cap in the photo — SAM 2 segments only the one that was pointed to, while SAM 3 finds and segments both at once.

Why It Matters

• More efficient data labeling: processing every instance of a concept with one prompt can meaningfully cut down dataset construction and labeling time.
• Pairs well with video tracking: combined with tracking, concept-based segmentation opens up applications like continuously tracking a category of object across a full video — sports analytics, in-store movement analysis, and similar use cases.
• Ecosystem integration: SAM 3 has been integrated into the Ultralytics package (version 8.3.237), making it relatively straightforward to adopt in pipelines already built around earlier SAM models.

See Also

Our article SAM vs. YOLO: Comparing Segmentation and Detection covers the fundamental difference between the SAM family and YOLO — detection versus segmentation. SAM 3 can be read as an extension of the “segmentation” side of that comparison, now operating at the level of concepts.

Exact benchmark numbers, licensing terms, and API details may continue to change, so check Meta AI Research’s official announcement directly before adopting it in production.