- Meta released SAM 3.1 on March 27, 2026, introducing Object Multiplex for up to 7x faster multi-object tracking on a single H100 GPU with zero accuracy loss.
- SAM 3 introduced open-vocabulary concept segmentation, reaching 75-80% of human performance across 270,000 unique concepts — 50x more than existing benchmarks.
- The model uses an 848M parameter architecture combining a DETR-based detector with SAM 2’s transformer encoder-decoder tracker, sharing a single vision encoder.
What Happened
Meta Superintelligence Labs released SAM 3.1, an update to the Segment Anything Model 3 that delivers up to 7x faster inference for multi-object tracking in video. The update, published on March 27, 2026, introduces a technique called Object Multiplex that processes multiple objects in a single forward pass, eliminating the per-object bottleneck that limited SAM 3’s video tracking speed.
SAM 3, the base model released in November 2025, had already represented a significant architectural shift from its predecessors. It introduced open-vocabulary concept segmentation, allowing users to specify what to segment using text phrases or image exemplars rather than manual point-and-click prompts. SAM 3.1 preserves all of that capability while making it fast enough for real-time production use.
Why It Matters
Speed has been the primary practical constraint for deploying segmentation models in production environments. SAM 3 could identify and segment objects accurately, but tracking dozens of objects simultaneously in video was computationally expensive and too slow for applications that require real-time processing.
SAM 3.1 addresses this directly. At 128 objects on a single H100 GPU, SAM 3.1 runs 7x faster than the November 2025 release. For medium object-count videos, the improvement translates from 16 to 32 frames per second, crossing the threshold for real-time applications in robotics, video editing, and autonomous systems. Importantly, the speedup comes with zero loss in segmentation accuracy.
The open-vocabulary capability in SAM 3 expanded what segmentation models can understand. The system reaches 75-80% of human performance on Meta’s SA-CO benchmark, which contains 270,000 unique concepts. That is over 50 times more concepts than existing segmentation benchmarks cover, meaning the model can segment objects that previous systems had never been trained to recognize.
Technical Details
SAM 3 uses an 848-million parameter architecture built around two components that share a single vision encoder. The detector is a DETR-based model conditioned on text, geometry, and image exemplars. The tracker inherits SAM 2’s transformer encoder-decoder architecture for maintaining temporal consistency across video frames.
A key architectural addition is the “presence token,” which improves discrimination between closely related text prompts. This helps the model distinguish between similar concepts — for example, differentiating “golden retriever” from “labrador retriever” in the same scene, a task that tripped up earlier open-vocabulary models.
The training data behind SAM 3 comes from an automated data engine that annotated over 4 million unique concepts, creating what Meta describes as the largest high-quality open-vocabulary segmentation dataset to date. SAM 3.1’s Object Multiplex technique modifies the inference pipeline without changing the underlying model architecture or weights. Existing SAM 3 installations can be updated via a code pull and checkpoint download rather than full retraining.
Who’s Affected
Robotics and autonomous systems developers benefit most from the real-time tracking capability. Applications that require segmenting and tracking multiple objects simultaneously — warehouse automation, surgical robotics, autonomous driving, drone navigation — can now use SAM 3.1 without requiring multi-GPU setups for acceptable frame rates.
Video editing and visual effects teams gain a tool that can segment any describable concept at production-relevant speeds. The open-vocabulary approach removes the need for custom training on specific object categories, reducing setup time for new projects.
Researchers working on downstream computer vision tasks can access the model through Hugging Face with authentication. The development team includes more than 40 researchers at Meta Superintelligence Labs.
What’s Next
SAM 3.1 is available under the SAM License on GitHub and Hugging Face. System requirements include Python 3.12+, PyTorch 2.7+, and CUDA 12.6+, which limits deployment to relatively recent GPU hardware and rules out older workstation setups.
The 75-80% human performance figure on concept recognition leaves room for continued improvement. Whether Meta will release lighter model variants optimized for edge devices and mobile deployment remains an open question that would broaden SAM 3’s practical reach beyond data center environments.
