SAM 3 is Meta’s Segment Anything Model 3, an open-source foundation model for image and video segmentation with more than 12,000 stars on GitHub, providing code for inference, fine-tuning, and downloading the trained checkpoints.
The Segment Anything family changed computer vision by making segmentation promptable: instead of training a model per category, you point at what you want — with a click, a box, or an example — and the model returns a precise mask. SAM 3 pushes that idea further toward concept-level understanding, segmenting objects described by short text phrases and visual prompts across both images and video. The repository is the complete working toolkit: inference code to run the model, fine-tuning code to adapt it to specialized domains, links to the official trained checkpoints, and example notebooks that walk through real usage end to end. As with its predecessors, Meta releases SAM as research infrastructure for the whole field — the models power annotation tools, editing products, robotics pipelines, and scientific imaging workflows far beyond Meta itself. If SAM 2 brought promptable segmentation to video, SAM 3 is the generation that makes open-vocabulary, concept-driven segmentation practical.
SAM 3 is free: the code and trained checkpoints are published by Meta for the community to download and use, subject to the license terms included in the repository. Your only cost is the compute you run it on.
Computer vision engineers and researchers, annotation and labeling teams, video editors, and anyone building products that need fast, precise object masks without training a segmentation model from scratch.