SAM 3 is Meta's unified foundation model for promptable segmentation across images and videos, released by Meta Superintelligence Labs. The repository provides code for running inference and finetuning with the model, along with trained model checkpoints and example notebooks demonstrating usage. The model represents a significant advancement over its predecessor SAM 2 by introducing the ability to exhaustively segment all instances of open-vocabulary concepts specified through short text phrases or visual exemplars. Unlike prior work, SAM 3 handles a vastly larger set of open-vocabulary prompts and achieves 75-80% of human performance on the new SA-CO benchmark, which contains 270,000 unique concepts—over 50 times more than existing benchmarks.

The breakthrough underlying SAM 3 stems from an innovative data engine that has automatically annotated over 4 million unique concepts, creating the largest high-quality open-vocabulary segmentation dataset to date. The model introduces a new architecture featuring a presence token that improves discrimination between closely related text prompts, such as distinguishing "a player in white" from "a player in red." The architecture employs a decoupled detector-tracker design that minimizes task interference and scales efficiently with data. SAM 3 consists of 848 million parameters, with a detector based on DETR conditioned on text, geometry, and image exemplars, while the tracker inherits the SAM 2 transformer encoder-decoder architecture to support video segmentation and interactive refinement.

The repository is written in Python and includes multiple example notebooks covering different use cases. The sam3_image_predictor_example notebook demonstrates prompting with text and visual box prompts on images, while sam3_video_predictor_example shows text prompting on videos with interactive refinement using points. Additional notebooks cover batched inference on images, the SAM 3 Agent for complex text prompts, and visualization of examples from the SA-Co evaluation sets. The examples directory contains notebooks for interactive instance segmentation in both images and videos, integration with multimodal large language models, and evaluation procedures on the SA-Co dataset.

Installation requires Python 3.12 or higher, PyTorch 2.7 or higher, and a CUDA-compatible GPU with CUDA 12.6 or higher. Users must request access to model checkpoints on the SAM 3 Hugging Face repository and authenticate before downloading. A significant update occurred on March 27, 2026, when SAM 3.1 Object Multiplex was released, introducing a shared-memory approach for joint multi-object tracking that improves speed without sacrificing accuracy. This update includes improved model checkpoints available on Hugging Face, requiring users to pull the latest code and reinstall the repository.

According to GitGenius activity tracking, the repository shows median issue and pull request response latency of 14.7 hours across 259 items, with a mean latency of 130.6 hours. The most active contributors tracked by GitGenius are heyoeyo with 68 events, alcinos with 60 events, and keejkrej with 22 events. The repository connects to related projects through overlapping contributors with facebookresearch/sam-3d-objects, qwenlm/qwen3-vl, and huggingface/transformers. GitGenius classifies the repository across multiple domains including 3D segmentation, point clouds, meshes, volumetric data, computer vision, deep learning, foundation models, interactive segmentation, automatic segmentation, and prompting techniques.