Meta Segment Anything (SAM 3)
Meta's open-weight Segment Anything models: SAM 3 and 3.1 find, segment and track every instance of a concept in images and video from text or clicks
Overview
Segment Anything (SAM) is Meta's family of open-weight models that cut out objects in images and video. The current version, SAM 3, released on November 19, 2025, adds what Meta calls promptable concept segmentation: give it a short noun phrase or an example image, and it finds, masks and tracks every matching instance, while still taking the classic point, box and mask prompts. SAM 3.1, released March 27, 2026, is a drop-in replacement with faster multi-object video tracking. Weights are on Hugging Face (access to facebook/sam3 must be requested) and you can try it in Meta's Segment Anything Playground. The earlier SAM 2 remains available under Apache 2.0 and segments one object per prompt; SAM 3 uses Meta's own SAM License.
The Verdict
Who Should Use Segment Anything?
Best For
- Vision engineers who need open-vocabulary segmentation without training
- Labeling pipelines that pre-annotate images and video
- Video tracking of many objects from a text prompt
- Teams that want to fine-tune a strong base model on a small dataset
Not Ideal For
- Projects that need a standard permissive license (use SAM 2 instead)
- Fine-grained specialist terms out of the box, such as medical cell types
- Long, complex referring phrases without pairing it with a multimodal LLM
What's Great
- Text and image-exemplar prompts find all instances of a concept
- Meta reports a 2x gain over prior systems on its SA-Co benchmark
- About 30 ms per image with 100+ objects on an H200, per Meta
- Free weights, code and a large new dataset
- Independent testers found local video mask propagation works well
Watch Out For
- Custom SAM License, not Apache; it bans military and some other uses
- Weak zero-shot results on fine-grained out-of-domain concepts
- Struggles with long referring expressions
- Video cost grows with the number of tracked objects
- Needs Python 3.12+ and PyTorch 2.7+; Hugging Face weights are gated
Pricing
View all features & details
Key Features
- Promptable concept segmentation from noun phrases
- Image-exemplar, point, box and mask prompts
- Video detection and tracking with interactive refinement
- SAM 3.1 Object Multiplex for faster multi-object video
- SA-Co benchmark and datasets on Hugging Face
- Companion SAM 3D models for 3D reconstruction
Use Cases
- Auto-labeling training data
- Video editing and effects masks
- Counting and tracking objects in footage
How It Compares
| Feature | SAM 3 | SAM 2 | Ultralytics YOLO |
|---|---|---|---|
| Text prompts | Yes | No | See vendor |
| Instances per prompt | All matching | One | Fixed classes |
| License | SAM License | Apache 2.0 | See vendor |
| Video tracking | Yes | Yes | See vendor |
Sources: Meta SAM 3 blog · SAM 3 repo · SAM 2 repo · SAM 3.1 on Hugging Face
Is this your tool? Add a badge