SAM 3 Is Here: Meta's Latest Vision AI Can Now Understand Your Words

Source: Dev.to

SAM lineage

SAM 1 (2023)

• Zero‑shot model that segments any object using clicks or bounding boxes.

SAM 2 (2024)

• Added video support, enabling object tracking across frames.

SAM 3 (2025)

• Introduces native text‑prompt understanding and 3D reconstruction capabilities.

What’s new in SAM 3

• Text‑prompt segmentation – Describe what you want to segment (e.g., “red car”, “yellow school bus”, “impala”) and the model automatically detects, masks, and tracks the objects.
• Unified image‑video backbone – A shared vision encoder processes individual frames while preserving temporal consistency, eliminating the need for separate detection and tracking pipelines.
• 3D reconstruction (“SAM 3D”) – Estimates an object’s three‑dimensional shape from 2D images or video, opening possibilities for AR/VR, robotics, and XR applications.
• Optimized inference – Despite added functionality, SAM 3 remains efficient, outperforming earlier versions on Meta’s SA‑Co dataset while being designed for edge‑device deployment.

Technical deep‑dive: Local implementation on AMD Ryzen AI Max+ 395

Hardware configuration

• CPU: 16‑core Zen 5 (Strix Halo)
• Memory: 128 GB LPDDR5x (8000 MT/s)
• Peak performance: up to 126 TOPS

By leveraging the Ryzen AI’s unified memory architecture, SAM 3 runs locally without cloud dependencies, offering low latency and data‑privacy benefits. While high‑end GPUs (e.g., NVIDIA H100) are typical for large vision models, the Ryzen platform provides a cost‑effective alternative for workloads with modest memory footprints and real‑time requirements.

Running SAM 3 on the Ryzen AI Max+ 395 delivers impressively fast inference, enabling edge‑camera deployments for “segment‑by‑description” detection.

A full implementation guide—including code, benchmark results, and integration with IoT edge cameras—will be published in a forthcoming article.

Resources

• GitHub:
• Hugging Face:
• Demo:

Try the official demo to experience SAM 3’s accuracy firsthand. Your feedback is welcome in the comments.