// the find
facebookresearch/sam3
The repository provides code for running inference and finetuning with the Meta Segment Anything Model 3 (SAM 3), links for downloading the trained model checkpoints, and example notebooks that show how to use the model.
SAM 3 is Meta's follow-up to SAM 2 for promptable image and video segmentation, adding open-vocabulary concept detection so you can hand it a short text phrase (or an example box/point) and get every matching instance segmented and tracked. It's aimed at CV researchers and engineers building detection/tracking pipelines who need more than fixed-class detectors like YOLO or Grounding DINO.
The decoupled detector-tracker design sharing one vision encoder is a real architectural choice, not just a bigger model - it lets them scale detection and tracking data independently instead of one task starving the other. They back the open-vocabulary claim with an actual benchmark (SA-Co, 270K concepts) and published comparisons against DINO-X, OWLv2, and Gemini 2.5, not just internal numbers. The video predictor exposes a session-based API (start_session/add_prompt) that supports interactive point refinement mid-stream, which is the part people actually struggle to build themselves. It's also actively maintained - the 3.1 update landed with a legitimate efficiency win (shared-memory multi-object tracking) rather than a version bump for marketing.
Checkpoints are gated behind a Hugging Face access request, so it's not a clone-and-run repo - budget time for approval before you can test anything. The dependency floor is narrow: Python 3.12+, PyTorch 2.7+, CUDA 12.6+, and the fast path needs flash-attn-3 built from a specific wheel index plus a custom cc_torch fork - this will fight you on any existing environment. At 848M params with those attention/CUDA requirements, there's no realistic CPU or edge story here despite the API looking lightweight. Install is conda-first with pip extras split across notebooks/train/dev groups, so it's easy to pip install -e . and then hit missing-dependency errors the first time you open an example notebook.