Robots Atlas>ROBOTS ATLAS
SAM 2
AI Modelsโ€บSegment Anything (SAM)

SAM 2

2 (Segment Anything Model 2)ย ยทย Family: Segment Anything (SAM)
Meta's open (Apache 2.0) foundation model for real-time promptable object segmentation and tracking in images and video; Hiera encoder + streaming memory.
โœ“ Activeโœ“ Public accessโš– Open sourceVisionSpecialized AI๐Ÿ“ Segment Anything (SAM)
Parameters
Warianty: tiny/small/base+/large (Hiera). Sam2-hiera-large ~224M.
parameters
Release date
29 July 2024
๐ŸขMeta AIProducer
Access:DownloadAPIHostedDeployment:๐Ÿ’ป Localโ˜ Cloud

Overview

SAM 2 (Segment Anything Model 2) is a foundation model developed by Meta AI (FAIR), released on July 29, 2024. It is a unified model for real-time promptable object segmentation and tracking - in both images and video. It is the successor to SAM, which worked on images only.

The user indicates an object with a prompt (point, box or mask) on any frame, and the model generates its pixel mask and propagates it across the whole video sequence (segmentation + tracking). The architecture includes a Hiera image encoder (hierarchical Vision Transformer), a prompt encoder, a mask decoder and a streaming memory module. The model features zero-shot generalization to previously unseen objects.

SAM 2 is released under the permissive Apache 2.0 license together with the SA-V dataset (about 51k videos and over 600k masklets). The model is available via Hugging Face and the GitHub repository (facebookresearch/sam2), and on Amazon SageMaker JumpStart (as Segment Anything 2.1). It comes in several size variants (tiny/small/base/large).

Classification
VisionSpecialized AI
Family: Segment Anything (SAM)
Applications
Data curation & annotationVideo analyticsRobotic manipulationImage editing
Access & deployment
DownloadAPIHosted
LocalCloud
Weights: Open source
Key parameters
๐Ÿงฉ Parameters: Warianty: tiny/small/base+/large (Hiera). Sam2-hiera-large ~224M.
โœ“ Fine-tuning
๐Ÿ“ฅ Input: image, video, structured data
Platforms
Hugging Face Hub

Technical specification

Parameters
Warianty: tiny/small/base+/large (Hiera). Sam2-hiera-large ~224M.
parameters
License
Apache 2.0
Hardware requirements
Variants from tiny (mobile/edge) to large; the ViT (Hiera) encoder benefits from GPU. Supported: PyTorch, Transformers. Near-real-time video inference.
Features:โœ“ Fine-tuning
Modalities
โฌ‡ Input
imagevideostructured_data
โฌ† Output
imagestructured_data

Capabilities and applications

Native model capabilities
Image segmentation
The ability to isolate objects or regions in an image by generating pixel masks (e.g. based on points, boxes or other prompts).
Category: vision
Object tracking (video)
The ability to track selected objects across consecutive video frames, maintaining their masks/identity despite motion, occlusion and appearance changes.
Category: vision
Image understanding
Analysing and interpreting the content of images.
Category: vision
Video understanding
The model's ability to analyse and interpret video content โ€” recognising actions, motion, events and relationships between objects over time.
Category: video
Application domains
Data curation & annotationVideo analyticsRobotic manipulationImage editing

Pricing

Technical architecture

Deployment and security

โ˜ Available on platforms
โ˜Hugging Face HubPlatform
๐Ÿ”’ Security / Enterprise
โœ“ Verified enterprise information

Open-source model under Apache 2.0 - full freedom of self-hosting / on-premises with data control (local processing, no need to send images/video to the cloud).

Updated: 31 Jul 2026โ†— Security documentation

Sources and related pages