SAM 3
Segment anything — now including things you name
What it is
Meta's unified 848M-parameter foundation model for promptable segmentation in images and video: a DETR-based detector fused with the SAM 2 tracker, so one model detects, segments, and tracks from text prompts, points, boxes, or masks — with open-vocabulary concepts trained over ~270k unique terms.
Why it's interesting
The first SAM generation to do open-vocabulary segmentation from free text plus unified image-and-video tracking in one model. It anchors the entire segmentation ecosystem — but its weights sit under Meta's custom gated license, not an OSI one.
Use cases
- Video object tracking
- Open-vocabulary detection and segmentation
- Dataset auto-labeling
- Robotics perception
Who it's for
CV researchers, ML engineers, annotation teams
Setup
Moderate. Python 3.12+, PyTorch 2.7+, CUDA GPU; checkpoint download requires Hugging Face approval
Limitations & cautions
Non-OSI gated weights with use restrictions, and heavy GPU needs for real-time video.
Editorial takeaway
Technically indispensable, legally 'read the license first.' Both things are true, and this site exists to say so in the same breath.