# Meta updates SAM to 3.1 for faster video tracking > SAM 3.1 adds multiplexed object processing and global reasoning, with new artifacts for teams evaluating video detection and tracking throughput. Canonical URL: https://www.devobs.io/news/news-meta-sam-3-1-video-update/ By: Amara Okafor Published: 2026-09-06T11:58:54.632Z Updated: 2026-09-06T11:58:54.632Z Event date: 2026-03-27 Section: AI Meta announced the SAM 3.1 update on March 27, adding multiplexing and global reasoning to its video detection and tracking system. The dated [update](https://ai.meta.com/blog/segment-anything-model-3/) appears above the earlier SAM 3 announcement and links to refreshed model artifacts. ## Multiple objects share a processing pass According to Meta, SAM 3.1 can process up to 16 objects in one forward pass. The company reports an increase from 16 to 32 frames per second on an H100 for workloads with a medium number of objects. Updated checkpoints, code, a paper and the playground accompany the release. Those figures describe Meta’s reported test conditions. They should not be interpreted as a frame-rate guarantee for a different accelerator, video resolution or scene. ## Measure tracking quality alongside speed Our analysis: throughput is only one part of a usable video pipeline. An evaluation should include decoding, preprocessing and result delivery, rather than timing the model alone. Warm-up behavior and batching can also change the latency seen by an interactive application. Test crowded scenes, occlusion and objects leaving and re-entering the frame. A system can process frames quickly while making identity errors that undermine an editor, analytics workflow or annotation job. Record how often a user must repair a track, and whether those repairs remain consistent in later frames. Multiplexed processing also makes object count an important test variable. Measure sparse and dense scenes instead of assuming one benchmark represents both. The release gives teams a concrete new version to compare with their current setup. Pin the checkpoint and evaluation inputs, preserve the previous result set, and judge adoption on the combined cost of processing and correcting the output. ## Source references -