SAM 2.1

SAM 2.1 is Meta's refined follow-up to Segment Anything 2, with updated checkpoints released September 29, 2024. Like SAM 2 it is a unified model for promptable object segmentation in images and video, using a streaming memory architecture that processes video in real time and tracks objects across frames.

It ships in four sizes from tiny (38.9M parameters, 91 FPS) to large (224.4M parameters), and the 2.1 update specifically improves performance on visually similar objects and occlusions. The model family is open, with code and checkpoints published, and Meta later made SAM 2.1 available through Amazon SageMaker JumpStart in February 2025.

The honest caveat is the usual one for promptable segmentation: output quality still depends heavily on prompt quality, and dense or ambiguous scenes need careful prompting. Real-world uses reported by Meta range from video data annotation to wildlife tracking and disaster-relief imagery.