SAM 3
Promptable concept segmentation across images and video.
About
SAM 3 is Meta's November 2025 successor to the Segment Anything line. The big change is promptable concept segmentation: describe a concept with a short noun phrase or an example image and it returns masks plus stable identities for every matching instance at once, across images and video — earlier versions handled one object per prompt. Meta released the weights, the SA-Co open-vocabulary benchmark, fine-tuning code, and a companion SAM 3D suite.
It shines on common visual concepts but is weaker on rare ones outside its 4M-concept training set, and long, complex text prompts still trip it up. Ships under Meta's SAM License, a custom research-friendly license rather than a standard OSI one.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SA-Co benchmark: roughly 75-80% of human-level performance (third-party reporting of Meta's figures)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Meta release Nov 2025 with checkpoints, SA-Co benchmark, training/fine-tuning code, and SAM 3D suite; Concept-segmentation (prompt by noun phrase or example) is a genuine capability jump over SAM 2; Weights, eval harness, and demo code all public; training data (SA-Co's 4M concepts) described but not fully released; Honest limits: weak on rare concepts, struggles with long complex text prompts. License: SAM License (Meta custom, research-friendly) — not OSI-open.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Meta release Nov 2025 with checkpoints, SA-Co benchmark, training/fine-tuning code, and SAM 3D suite
- · Concept-segmentation (prompt by noun phrase or example) is a genuine capability jump over SAM 2
- · Weights, eval harness, and demo code all public; training data (SA-Co's 4M concepts) described but not fully released
- · Honest limits: weak on rare concepts, struggles with long complex text prompts