Meta's SAM 3.1 tracks 128 objects 7x faster
Meta just dropped SAM 3.1, adding a shared-memory approach called Object Multiplex to its visual segmentation model. It delivers roughly 7x faster inference when tracking 128 objects on a single H100 without dropping accuracy, beating prior scores on 6 out of 7 video benchmarks. The model checkpoints are available to request on Hugging Face.
SAM 3.1 is Meta's unified visual foundation model for promptable image and video segmentation. It detects, segments, and tracks objects using text prompts, points, bounding boxes, or masks.
How it gets faster
While SAM 3 already brought open-vocabulary concept segmentation covering over 50x more concepts than existing benchmarks, SAM 3.1 tackles scaling across many targets. The key addition is Object Multiplex, a shared-memory approach designed for joint multi-object tracking.
Running on a single H100 GPU, it achieves around 7x faster inference when tracking 128 objects simultaneously, without sacrificing accuracy. It also posts improved video object segmentation scores on 6 out of 7 standard benchmarks.
The model checkpoints are available on Hugging Face, gated behind a simple agreement to Meta's privacy policy. Keep in mind that this repository only hosts raw checkpoints rather than a native Hugging Face Transformers integration. You will need to grab the actual code, environment setup, and docs from the SAM 3 GitHub repo.
Why it matters
If you deal with crowded video tracking—like surveillance, sports analytics, or autonomous driving footage—multi-target compute has always been a bottleneck. Tracking a hundred distinct entities usually blows up VRAM or stalls frame rates.
We think Meta's shared-memory multiplexing is a neat systems win. Getting a 7x speedup on heavy scenes on a single card means real-time dense tracking is finally practical without stitching together massive clusters.
Sources (1)
- huggingface.co(opens in a new tab)Originalhuggingface.co