Meta Launches SAM 3 on GitHub: AI Model Detects and Tracks Objects Across 4 Million Concepts

Meta's SAM 3 is now live on GitHub, a groundbreaking AI model that detects, segments, and tracks objects across 4 million unique concepts in images and videos, achieving up to 80% of human performance — with an updated SAM 3.1 delivering even faster multi-object tracking via a new shared-memory approach.

• Meta's SAM 3 (Segment Anything with Concepts) is now available on GitHub, offering a unified foundation model capable of detecting, segmenting, and tracking objects in images and videos using text or visual prompts.
• SAM 3 achieves 75-80% of human performance on the new SA-CO benchmark — which spans 270,000 unique concepts — powered by a data engine that has automatically annotated over 4 million unique concepts, making it the largest high-quality open-vocabulary segmentation dataset to date.
• A SAM 3.1 update released on March 27, 2026 introduces Object Multiplex, a shared-memory approach for joint multi-object tracking that delivers significantly faster performance without sacrificing accuracy, with new model checkpoints available on Hugging Face.