SAM 3

SAM 3 (Segment Anything Model 3) is Meta's November 19, 2025 release of the Segment Anything family, introducing Promptable Concept Segmentation. Where SAM and SAM 2 needed points, boxes, or masks, SAM 3 accepts a short noun phrase such as "yellow school bus", an example image, or both, and returns unique masks and IDs for every matching instance at once across images and video, with tracking across frames. It ships alongside the SA-Co benchmark dataset: 120K images and 1.7K videos with exhaustive concept-level masks spanning over 200K unique concepts. Meta reports a 2x performance improvement over existing systems on SA-Co and roughly 30 milliseconds to process a single image with 100+ detected objects on an NVIDIA H200.

The weights are released through a gated Hugging Face repository (facebook/sam3, manual approval) with open-source inference and fine-tuning code under Meta's custom SAM License, which allows research and commercial use with restrictions — it is not an OSI-approved license. The technical report is arXiv 2511.16719. Benchmark and speed figures are Meta's own; the release date, license terms, gated availability, and concept-prompting capability are independently confirmable.