Meta official inference-efficiency plot
H100 FPS: 1 tracked object
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
SPIRITT
SAM 3.1SAM 3.1 tracks the crowd in one shared memory
Meta's drop-in SAM 3 update adds Object Multiplex, grouping tracked objects into shared-memory buckets instead of running a separate path for each one. Meta reports roughly seven times the throughput at 128 objects on one H100. SAM 3.1 is accessible through SPIRITT AI Gateway.
SAM 3.1 release and SAM 3 builder context
Official SAM 3.1 release posts and a multi-model C++ implementation, followed by clearly dated SAM 3 family demos and firsthand security-camera workflows. SAM 3 and EdgeTAM examples are context, not SAM 3.1 benchmark reruns.
Object Multiplex removes the per-object bottleneck
The update is primarily an inference-engine improvement: shared video memory, joint tracking, and optimized GPU execution, with capability gains that are real but not uniform.
What Meta changed, and where SAM 3.1 fits
SAM 3.1 is a drop-in checkpoint and code update for SAM 3's promptable image and video segmentation stack. It retains text, point, box, and mask prompts while adding Object Multiplex for joint multi-object tracking.
The original SAM 3 pipeline processed tracked objects independently, so compute scaled with object count. Object Multiplex groups objects into fixed-capacity buckets, shares memory, and processes up to 16 objects in one forward pass. Meta reports 16 to 32 FPS for a medium object count and roughly seven times SAM 3 throughput at 128 objects on a single H100.
Official release notes show stronger VOS results on six of seven benchmarks and gains on YT-Temporal-1B. They also show regressions on some public video detection metrics, including LVVIS, BURST, and YTVIS21, so faster does not mean universally more accurate.
SAM 3.1 is accessible through SPIRITT AI Gateway. Open a SPIRITT Workspace and ask the agent to use SAM 3.1 for the image or video task; the gateway handles model access while the workspace manages media, code, experiments, tools, and the surrounding automation.
Object MultiplexUp to 16 objects / pass~7x at 128 objectsText + visual promptsImage + videoMeta API list: $2.50 / 1K imagesMeta API list: $0.20 / 1K video framesGated checkpoint
Relative throughput at 128 tracked objects
WinnerMeta release notes on a single H100 GPU. The roughly seven-times result compares SAM 3.1 with the November 2025 SAM 3 release at 128 tracked objects; it is not a cross-vendor benchmark.
SAM 3.1 accuracy, throughput, and implementation evidence
All 17 accuracy comparisons and all eight object-count measurements from Meta's release, including regressions, plus separately labeled Apple Silicon implementation reports.
Meta official inference-efficiency plot
H100 FPS: 2 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 4 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 8 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 16 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 32 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 64 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Meta official inference-efficiency plot
H100 FPS: 128 tracked objects
Values as rounded on the release-note efficiency plot. H100, same plotted workload; speed depends on object count. Optimized SAM 3 is faster for the one-object case.
Community implementation: M3 Max bf16
MLX tracking: 3 objects (ms/frame)
mlx-community/sam3.1-bf16 model-card report. Apple M3 Max, bf16, implementation-specific and not comparable to Meta's H100 tests. Lower is better.
Community implementation: M3 Max bf16
MLX tracking: 4 objects (ms/frame)
mlx-community/sam3.1-bf16 model-card report. Apple M3 Max, bf16, implementation-specific and not comparable to Meta's H100 tests. Lower is better.
Community implementation: M3 Max bf16
MLX tracking: 5 objects (ms/frame)
mlx-community/sam3.1-bf16 model-card report. Apple M3 Max, bf16, implementation-specific and not comparable to Meta's H100 tests. Lower is better.
Community implementation: 224px
MLX optimized real-time (ms/frame)
MLX model-card report, M3 Max bf16 at 224px with backbone caching and tracker propagation. Approximate times, not original-resolution full detection every frame.
SA-Co/VEval: SA-V
SA-V cgF1
Benchmark test split. A 0.2-point gain.
SA-Co/VEval: SA-V
SA-V pHOTA
Benchmark test split. A 0.7-point gain.
SA-Co/VEval: YT-Temporal-1B
WinnerYT-Temporal cgF1
Benchmark test split. The largest reported concept-tracking gain at 2.1 points.
SA-Co/VEval: YT-Temporal-1B
YT-Temporal pHOTA
Benchmark test split. A 0.8-point gain.
SA-Co/VEval: SmartGlasses
SmartGlasses cgF1
Benchmark test split. SAM 3.1 is 0.1 points lower.
SA-Co/VEval: SmartGlasses
SmartGlasses pHOTA
Benchmark test split. A 0.8-point gain.
Public video PCS regression
LVVIS test mAP
Official release notes. SAM 3.1 is 2.0 points lower.
Public video PCS regression
BURST test HOTA
Official release notes. SAM 3.1 is 1.2 points lower.
Public video PCS regression
YTVIS21 val mAP
Official release notes. SAM 3.1 is 0.8 points lower.
Public video PCS
WinnerOVIS val mAP
Official release notes. A 1.0-point gain.
Video object segmentation
WinnerMOSEv1 J&F
Official release notes, validation split. A 1.2-point gain.
Video object segmentation
DAVIS17 J&F
Official release notes, validation split. A 0.5-point gain.
Video object segmentation
LVOSv2 J&F
Official release notes, validation split. A 0.7-point gain.
Video object segmentation
SA-V val J&F
Official release notes, validation split. A 0.3-point gain.
Video object segmentation
SA-V test J&F
Official release notes, test split. A 0.7-point gain.
Video object segmentation regression
YTVOS19 G
Official release notes, validation split. SAM 3.1 is 0.4 points lower.
Challenging video segmentation
WinnerMOSEv2 J&αΈ
Official release notes, validation split. A 2.0-point gain.
Meta's official release notes supply all 17 accuracy comparisons and eight H100 object-count conditions. The separate MLX model card supplies the labeled M3 Max implementation timings; these are not comparable to the H100 setup. Timing depends on object count, resolution, precision, caching, and compilation. Meta API list prices are developer-published references, not a SPIRITT billing quote. SAM 3 and EdgeTAM posts are family context rather than SAM 3.1 reruns.
How It Works
Ask the workspace agent to use SAM 3.1 through SPIRITT AI Gateway
01
Open a workspace
Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.
02
Just Ask to use SAM 3.1
Tell your workspace agent what to detect, segment, or track and provide the image or video. SPIRITT AI Gateway handles SAM 3.1 access, so the agent can focus on the vision workflow rather than provider setup.
03
Build or automate
Use the workspace terminal, files, code, browser, and persistent storage to ingest video, run segmentation, inspect masks, benchmark throughput, and connect outputs to the rest of your agentic application.
Turn masks into a working vision system.
Ask your SPIRITT Workspace to use SAM 3.1, attach the media, and let the agent connect masks and tracks to the rest of your application.
Questions
SAM 3.1 FAQ
01What is SAM 3.1?+
Meta's March 2026 update to Segment Anything Model 3. It keeps SAM 3's promptable image and video segmentation capabilities while adding Object Multiplex and optimized multi-object video inference.
02What does Object Multiplex do?+
It groups tracked objects into fixed-capacity buckets and processes them jointly through shared memory. Meta says this removes redundant per-object compute, supports up to 16 objects per forward pass, and improves throughput as object count rises.
03How much faster is SAM 3.1?+
Meta reports video throughput doubling from 16 to 32 FPS for a medium number of objects on one H100 and roughly seven times the SAM 3 throughput at 128 tracked objects. Results on other hardware, video sizes, and precision settings will differ.
04Is SAM 3.1 more accurate everywhere?+
No. The official release reports stronger VOS performance on six of seven benchmarks and some concept-tracking gains, but also regressions on metrics such as YTVOS19 G, LVVIS mAP, and BURST HOTA.
05How can developers access SAM 3.1?+
SAM 3.1 is accessible in SPIRITT through SPIRITT AI Gateway. Meta also offers a direct API route and a gated Hugging Face checkpoint for teams that want to operate the model outside the managed gateway.
06What prompts and media does it support?+
SAM 3.1 inherits SAM 3's image and video workflows, including text concepts plus point, box, and mask prompts. It is a segmentation and tracking system, not a general text-generation model.
07How do I use SAM 3.1 in SPIRITT?+
Open a SPIRITT Workspace, provide the image or video, and ask the agent to use SAM 3.1. It is accessible through SPIRITT AI Gateway, so the workspace can route the vision step without separate provider setup.
Buy from builders who use what they sellBuilt using