What is SAM 3 (Segment Anything Model 3)?
SAM 3 is the third Segment Anything Model from Meta, released on 19 November 2025 with weights, code, fine-tuning scripts and a new benchmark. Earlier versions outlined one object at a time from a click or a box. SAM 3 adds what Meta calls promptable concept segmentation: give it a short noun phrase such as "a hardcover book" or an example box, and it returns a mask, box and score for every matching object in an image, and tracks them through a video.
It has 848 million parameters split between a DETR-style detector conditioned on text, boxes and example images, and a tracker inherited from SAM 2, sharing one vision encoder. A "presence token" first decides whether the concept is in the frame at all, which helps with near-miss phrases like "a player in white" against "a player in red". Meta built the training set with a human-and-AI data engine that labelled more than 4 million unique concepts.
On 27 March 2026 Meta released SAM 3.1 as a drop-in replacement, with Object Multiplex tracking up to 16 objects in one forward pass. SAM 3 runs inside Instagram's Edits app and Facebook Marketplace's View in Room feature. The weights are under Meta's own SAM License, which replaces the Apache 2.0 licence of SAM and SAM 2.
Is SAM 3 (Segment Anything Model 3) worth it?
SAM 3 is the open model to beat for finding and outlining things by name. On Meta's SA-Co/Gold image benchmark it scores a cgF1 of 54.1 for instance masks, against 24.6 for OWLv2, 21.3 for DINO-X and 13.0 for Gemini 2.5, and on box detection it edges DINO-X on COCO, 53.6 AP to 52.4. On LVIS masks it reaches 48.5 AP against OWLv2's 43.4. SA-Co is Meta's own test set of 270,000 concepts, so its largest margins come on home ground; the LVIS and COCO results are on public benchmarks.
The gap to people is still wide. Human annotators score 72.8 on the same SA-Co/Gold masks, so SAM 3 sits at roughly three-quarters of human level, which is Meta's own description. Video is harder: on the SA-V test it reaches 30.3 cgF1 against a human 53.1, and on SmartGlasses footage 36.4 against 58.5. For production labelling it is a strong first pass, not a replacement for review.
Its prompts are deliberately simple. Meta says SAM 3 handles short phrases like "a hardcover book" but not "the second to last book from the right on the top shelf"; longer or reasoning-heavy requests need SAM 3 Agent, a pattern where a multimodal language model breaks the request into short phrases and checks the returned masks. In video, each tracked object cost a separate pass in the original release, keeping it near real-time only for about five objects. SAM 3.1 addresses that: Meta reports throughput doubling from 16 to 32 frames a second on one H100 for a medium number of objects and about a 7 times speed-up at 128 objects.
The licence is the main step back. SAM and SAM 2 were Apache 2.0; SAM 3 needs a gated sign-up with name, date of birth and affiliation, bars military and ITAR-related uses and reverse engineering, and requires research papers to credit it. Commercial use is allowed, and nearly 1.9 million downloads in the past 30 days on Hugging Face suggest few have been put off.
Where it earns its place
An ML engineer at a warehouse-robotics company building training data
Label every pallet, forklift and person across thousands of hours of video
Prompt SAM 3.1 with one short phrase per class and let the tracker follow each object through the clip, then have annotators correct the masks. Meta's own figures put video accuracy well below human level, so budget for review rather than accepting raw output.
A computer-vision lead at an agriculture start-up spotting crop disease
Detect a niche concept that zero-shot prompts miss
Fine-tune from the base checkpoint with Meta's code or in Roboflow, which trains on an instance-segmentation dataset at 1008 by 1008 pixels. Roboflow advises fine-tuning only where zero-shot prompts fail on domain-specific classes.
A developer at a photo-editing app with 20 engineers
Let users select "the sky" or "all the people" by typing it
Run SAM 3 server-side to return a mask per matching object in about 30 ms per image on an H200, then pass the mask to the app's own editing tools; SAM 3 outlines but does not alter pixels. Check the SAM License's terms with legal before shipping, since it is no longer Apache 2.0.
Before you commit
- Setup
- Request access on Hugging Face and log in with a token, then install the GitHub package on Python 3.12+, PyTorch 2.7+ and CUDA 12.6+; Transformers also loads it. Notebooks cover image and video prompting, batched inference and SAM 3 Agent.
- Integrations
- Hugging Face Transformers, Meta's Segment Anything Playground, and Roboflow for annotation, fine-tuning, hosted inference and auto-labelling. SAM 3 Agent connects it to a multimodal language model for complex prompts.
- Data and privacy
- Self-hosted, images and video stay on your hardware. The SAM License allows commercial use but forbids military, ITAR-related and other trade-controlled uses and reverse engineering, and gating collects name, date of birth, country and affiliation under Meta's privacy policy.
- Cost at scale
- At 848M parameters one data-centre GPU serves it; Meta reports 30 ms per image with more than 100 objects on an H200, and SAM 3.1 at 32 frames a second on one H100 for a medium number of objects. Roboflow's hosted pricing follows its own plans.
- Limits
- Prompts must be short noun phrases or example boxes; in the original release video latency rises linearly with the number of tracked objects. Meta's SA-Co benchmark results are self-reported, and human annotators still score about 19 points higher on images.
What’s good
- 54.1 cgF1 on SA-Co/Gold, more than double OWLv2 and DINO-X (Meta)
- Detects, segments and tracks every instance of a text phrase
- 30 ms per image with 100+ objects on an H200 (Meta)
- Fine-tuning code released; Roboflow supports training on custom data
What’s not
- Roughly three-quarters of human accuracy; video well below
- Short noun phrases only; complex queries need an extra language model
- SAM License replaces Apache 2.0, with a gated sign-up and use restrictions
- Needs a CUDA GPU; video cost grows with object count before SAM 3.1
Best for
Computer-vision teams that need masks or boxes for named objects across images and video without training a detector for each class.
SAM 3 (Segment Anything Model 3) alternatives
SAM 3 (Segment Anything Model 3): common questions
- Is SAM 3 (Segment Anything Model 3) worth it?
- SAM 3 turns "outline every forklift in this video" into one prompt, at more than twice the accuracy of the open-vocabulary detectors before it, though still well short of a human labeller. We rate it 4.4 out of 5.
- How much does SAM 3 (Segment Anything Model 3) cost?
- Free under the SAM License after a gated sign-up on Hugging Face; commercial use allowed with military, ITAR and reverse-engineering restrictions. Hosted use through Roboflow is billed on Roboflow's plans.
- What are the downsides of SAM 3 (Segment Anything Model 3)?
- Roughly three-quarters of human accuracy; video well below; short noun phrases only; complex queries need an extra language model; sAM License replaces Apache 2.0, with a gated sign-up and use restrictions; needs a CUDA GPU; video cost grows with object count before SAM 3.1.
- Who is SAM 3 (Segment Anything Model 3) best for?
- Computer-vision teams that need masks or boxes for named objects across images and video without training a detector for each class.
- How much does SAM 3 (Segment Anything Model 3) cost for a team or in production?
- At 848M parameters one data-centre GPU serves it; Meta reports 30 ms per image with more than 100 objects on an H200, and SAM 3.1 at 32 frames a second on one H100 for a medium number of objects. Roboflow's hosted pricing follows its own plans.
- Is SAM 3 (Segment Anything Model 3) safe for company data?
- Self-hosted, images and video stay on your hardware. The SAM License allows commercial use but forbids military, ITAR-related and other trade-controlled uses and reverse engineering, and gating collects name, date of birth, country and affiliation under Meta's privacy policy.
- What are the best SAM 3 (Segment Anything Model 3) alternatives?
- The closest alternatives we have reviewed are Gemini 3.8 Flash and MiniCPM-V 4.7 (35B-A3B and 1B).
Answers checked Nov 2, 2026.
Make SAM 3 (Segment Anything Model 3)? Show this rating on your site
