Not runnable todaySegmentationVerified 2026-11-05
SAM 3 (Segment Anything with Concepts)
facebook/sam3
SAM 3 (Segment Anything with Concepts) is a image, video and text segmentation model (SAM License (Meta, 2025-11-19)). Not runnable on Mixpeek today.
Something blocks it today, such as the license, remote code or a GPU-only runtime. The reason is below.
No demo for SAM 3 (Segment Anything with Concepts) yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Segments and tracks every instance of a concept named by a short noun phrase or an exemplar image ('yellow school bus'), across images and video.
- Worst at
- About 0.86B parameters: GPU in practice; long or compositional phrases; access gated and approvals reported as rejected.
- Task
- Segmentation
- Input
- Image, Video, Text
- Output
- Masks, Boxes, Scores, Tracks
- Use cases
- Segmentation, Tracking, Object detection
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- image
- resolution
- 1008x1008
- video
- frames
- streamed with a tracker memory
- prompts
- short noun phrase, image exemplars (boxes), or SAM 2 style points/boxes
Output
- masks
- instance masks for every match
- boxes
- xyxy
- scores
- presence and instance scores
- tracks
- instance ids across frames
How Mixpeek wires it
- Extractors
- None today
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- None
- Combines with
- taxonomy labels -> SAM 3 concept prompts -> instance masks -> masked crops -> SigLIP
- SAM 3 instance counts per frame -> payload fields -> attribute filter
How it runs
- Status
- Not runnable today Blocked by gated hub access (manual approval, not guaranteed) and GPU-only practicality. Weights are safetensors with no remote code, so once access is granted it runs as a GPU plugin through transformers Sam3Model.
- Serve apps
- None
- Runtime
- transformers Sam3Model + Sam3Processor
- Weights
- safetensors, no remote code
- Parameters
- 860 M
- Size on disk
- 3,440 MB
- Preprocessing
- Resize to 1008, normalize; tokenize the concept phrase.
- Postprocessing
- Score threshold on presence and instance scores, mask upsampling.
- Serving
- both
- Hardware
- Needs a GPU
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 1
- Idle timeout
- 600 s
- Memory
- 12 GB
- CPUs
- 4
- GPUs
- 1
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed measurement; plan row 3 measures it in the standalone image.
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is SAM 3 (Segment Anything with Concepts) best at?
- Segments and tracks every instance of a concept named by a short noun phrase or an exemplar image ('yellow school bus'), across images and video.
- What are the limits of SAM 3 (Segment Anything with Concepts)?
- About 0.86B parameters: GPU in practice; long or compositional phrases; access gated and approvals reported as rejected.
- Can SAM 3 (Segment Anything with Concepts) be used commercially?
- Yes. Custom Meta license: royalty-free commercial use allowed, with carve-outs (military, nuclear, espionage, sanctioned parties, ITAR); redistribution must carry the same terms; Meta may change terms. Hub access is gated by manual approval.
- How does SAM 3 (Segment Anything with Concepts) run on Mixpeek?
- Blocked by gated hub access (manual approval, not guaranteed) and GPU-only practicality. Weights are safetensors with no remote code, so once access is granted it runs as a GPU plugin through transformers Sam3Model.