Run SAM 3 Weights Locally

Segment Anything 3 (SAM 3) is Meta's advanced model for image and video segmentation, capable of detecting, segmenting, and tracking objects using text prompts, clicks, bounding boxes, and image exemplars. Released in November 2025, SAM 3 can be easily run on various devices using the Roboflow inference Python package, which simplifies the setup process by automatically configuring hardware and optimizing runtime without requiring manual adjustments or installations. This package supports the use of SAM 3 across platforms, provides automatic updates, and offers a simple API for initiating models and obtaining segmentation masks. It integrates with the Roboflow ecosystem, facilitating tasks like auto-labeling datasets and visualization with the supervision library. SAM 3's ability to segment objects based on text prompts, known as Promptable Concept Segmentation, is highlighted along with its interactive segmentation workflow, which allows users to generate segmentation masks through human-in-the-loop annotation tools. The package also supports multiple output formats, ensuring flexible integration into various use cases. Additionally, SAM 3 can be used for auto-labeling datasets to train smaller models, offering a workflow to label images and create datasets suitable for real-time edge hardware applications. For those preferring to work with raw model weights, the option to download SAM 3's open-sourced codebase and checkpoints is available, albeit requiring more complex setup and management of dependencies.