Meta's SAM 3.1 Tracks 128 Objects 7x Faster on a Single GPU
Meta's SAM 3.1 adds Object Multiplex, a shared-memory tracking design that hits 7x faster inference on 128 objects without losing accuracy.
Takeaways
- Meta released SAM 3.1 as a drop-in upgrade to SAM 3 with new checkpoints on Hugging Face.
- Object Multiplex delivers ~7x faster inference at 128 objects on a single H100 GPU.
- Improved VOS scores on 6 of 7 benchmarks, including +2.0 on MOSEv2 and +2.1 cgF1 on YT-Temporal-1B.
- Optimizations include reduced CPU-GPU sync, better torch.compile fusion, and batched vision encoding.
- Concept segmentation capability from SAM 3 is unchanged, so no retraining or API changes required.
- Full technical details in Appendix H of the SAM 3 paper and release notes.
Meta released SAM 3.1 on March 27, 2026, as a checkpoint and inference update for large-scale video segmentation. Its new Object Multiplex pathway batches object-specific tracking work, reducing the serial processing that slowed dense scenes.
SAM 3 is Meta’s promptable segmentation foundation model. It accepts text, exemplar, click, and box prompts, then segments and tracks matching objects. Meta says its open-vocabulary system handles more than 50 times as many concepts as earlier benchmarks cover. A prompt such as “hands,” for example, can identify and track every matching instance in a video. SAM 3.1 preserves those capabilities while changing how the tracker processes multiple objects.
One pass across the object axis
SAM 3’s original video pipeline shared frame-level visual features but processed each object separately through memory attention and mask decoding. Compute therefore grew roughly in proportion to the number of tracked objects, making crowded scenes expensive.
Object Multiplex places object tracks into fixed-capacity groups and processes each group jointly while preserving separate identities and masks. Additional groups accommodate videos whose object count exceeds the configured capacity.