A segmentation model released by Meta on November 19, 2025. Give it a prompt describing what you want to cut out of an image or video, and it detects the matching objects and returns masks (the region data used for cutouts). SAM 1 and SAM 2 selected a single object per click or bounding box; SAM 3 also accepts short noun phrases such as “yellow school bus” or “striped cat”, or an example image, and picks up every matching object in the frame at once. Because it handles “open-vocabulary” concepts rather than a fixed label list, it works on targets that were never anticipated during training. Weights and code are published free of charge under Meta’s own SAM License.
Key Features
- Concept prompts in plain text: Pass a short noun phrase such as “solar panels” or “vehicles” and the model detects and masks every match. Meta calls this Promptable Concept Segmentation (PCS)
- Four prompt types you can combine: Besides text, you can use clicks (points), bounding boxes, masks, and image exemplars that say “find things like this object”, and combine them to narrow results
- Detects every instance at once: Where SAM 1 and SAM 2 returned one object per prompt, SAM 3 returns a separate mask and ID for each matching object
- Video tracking built in: Using the SAM 2-style masklet approach (masks that follow the same object across frames), it keeps tracking targets through a video
- Fast inference: Meta reports roughly 30 milliseconds per image on an H200 GPU, handling images with more than 100 objects
- Trained on the large SA-Co dataset: Meta built a concept-segmentation dataset covering more than four million noun phrases, and reports about a 2x gain over existing systems on its benchmark
Pricing
| Plan | Price | What you get |
|---|---|---|
| Weights and code | $0 | Published on GitHub (facebookresearch/sam3) and Hugging Face, subject to Meta’s SAM License |
| Segment Anything Playground | $0 | The official browser demo; upload images or video and check behavior without any technical setup |
Pricing is current as of August 2026. The SAM License differs from typical open source licenses, so always read the license text to confirm whether and how you may use it commercially. Check the official site for the latest information.
Pros & Cons
✅ Pros
- Because targets are named in text, you avoid collecting training data and retraining for each new class
- It works on unanticipated targets as-is, which suits settings where the objects are not fixed (inventory photos, field survey images, and so on)
- It detects all matching objects together, so images crowded with targets can be processed in one pass
- Weights and code are free to download, so you can run it on your own server or locally
- Video tracking is handled by the same model
⚠️ Cons
- A GPU is effectively required for practical speed, and the published figures assume high-end hardware such as an H200; CPU-only environments will struggle
- In video, processing time grows with the number of objects tracked at once, so following many targets in real time is demanding
- The SAM License is Meta’s own conditional license and does not offer the same freedom as MIT or Apache-2.0
- Text prompts work best as short noun phrases; complex sentence-level conditions are not its strength
- It is a model, not a platform ─ annotation management and dataset operations are still yours to build
Comparison with Similar Services
| Item | SAM 3 | SAM 2 | Grounded SAM | OWLv2 |
|---|---|---|---|---|
| Provider | Meta | Meta | Community (IDEA Research and others) | Google Research |
| Text prompts | Yes (short noun phrases) | No | Yes (via Grounding DINO) | Yes |
| Output | Masks (multiple instances) | Masks (one target per prompt) | Boxes plus masks | Boxes |
| Video | Yes | Yes | Depends on implementation | Mainly still images |
| Structure | Single self-contained model | Single model | Two models combined | Single model |
Who Is It For
- E-commerce and catalog teams that want to automate cutting products or parts out of photos
- Engineers whose target classes keep growing and who would rather not retrain every time the scope widens
- Video production and analysis teams that need to track a subject through footage and extract masks
- Data teams that want a model to do the first pass of annotation so people can focus on review and correction
- Research and R&D groups that want to try it at no cost first and see whether it holds up on their own data
Summary
SAM 3 widens the SAM family from “click to pick one thing” into “describe it in words and get everything that matches”. Not having to fix a label list pays off most in settings where the objects keep changing. On the other hand, the GPU-based running cost and the need to check Meta’s own license terms cannot be avoided. A good starting point is to drop your own material into the Segment Anything Playground and see whether a single prompt captures the targets you want. Note also that a speed-optimized SAM 3.1 was released in March 2026, worth considering when throughput matters.