SAM 3.1: Faster and More Accessible Real-Time Video Detection and Tracking With Multiplexing and Global Reasoning

Meta is treating its Segment Anything Model line like a piece of infrastructure that keeps getting tuned rather than a one-and-done release, and SAM 3.1 is the latest proof. The company says it spent the months since SAM 3 launched watching adoption climb and quietly reworking how the model chews through video, and the fix it landed on is something it calls object multiplexing.

The old way was blunt: every object being tracked got its own dedicated pass through the model. Track ten things, run ten passes. SAM 3.1 throws that out and processes everything together in a single forward pass, handling up to 16 objects at once. Meta frames this as global reasoning, and the practical payoff is concrete — on a single H100 GPU, throughput jumps from 16 to 32 frames per second for videos with a medium number of objects. That is a straight doubling of speed, not a marginal tweak.

What makes this matter beyond a benchmark chart is where the savings show up. Fewer redundant passes means less memory pressure and lower GPU demand overall, which Meta says opens the door to real-time tracking in crowded, complex scenes running on smaller, more accessible hardware rather than requiring a rack of high-end chips. It's also pitched as a drop-in replacement for SAM 3, so nobody building on the existing pipeline has to redo their integration to get the speed bump.

This lands on top of a busy stretch for the Segment Anything family. SAM 3 itself introduced promptable concept segmentation, letting people find and track objects using open-ended text or example images instead of a fixed label list, and it's already powering things like Facebook Marketplace's View in Room feature and an open wildlife-monitoring video dataset built with Conservation X Labs and Osa Conservation. SAM 3.1 doesn't change what the model can recognize — it changes how efficiently it does the recognizing when the video gets busy with multiple moving subjects, which is exactly the scenario that used to choke throughput.

Meta is pointing developers toward the SAM 3.1 checkpoint, the updated codebase and paper, and the Segment Anything Playground for anyone who wants to poke at it directly rather than take the throughput numbers on faith.