SAM 3 marks a major leap in visual AI by enabling concept-based segmentation—using simple phrases like 'yellow school bus' to detect and track every instance across media. It operates in just 30ms per image on an H200 GPU and scales efficiently for real-time video. A new benchmark, SACO, captures over 200,000 unique concepts, far surpassing prior limits, while an AI-powered data engine slashes annotation time from minutes to seconds using Llama-fine-tuned verifiers. The model introduces key innovations: a presence token to separate object recognition from localization, and a decoupled detector-tracker design for accurate video identity preservation. Integrated with multimodal LLMs like Gemini, SAM 3 acts as a visual reasoning tool, solving tasks such as distinguishing gender or comparing sizes. Fine-tuning with minimal examples enables adaptation to domains like medical imaging or autonomous driving. Through Roboflow, over 106 million smart polygons have been generated, saving more than 130 years of manual labeling across critical applications including cancer research and environmental cleanup.