The file is the SAM License, not Apache. Meta posted SAM 3.1 on 27 March 2026 as a drop-in video-tracking update to Segment Anything Model 3. Operator facts: a new checkpoint, a gated Hugging Face host, and Meta's own H100 speed table.

What it does

The facebookresearch/sam3 README describes SAM 3 as a unified model for promptable segmentation in images and videos using text or visual prompts (points, boxes, masks), including exhaustive segmentation of an open-vocabulary concept from a short phrase or exemplars. The 27 March 2026 latest-updates block says SAM 3.1 Object Multiplex "introduces a shared-memory approach for joint multi-object tracking that is significantly faster without sacrificing accuracy," with new checkpoints on Hugging Face and a requirement to pull latest repo code and reinstall.

RELEASE_SAM3p1.md says SAM 3's video pipeline processed each tracked object independently, scaling linearly, while Object Multiplex groups objects into fixed-capacity buckets and processes them jointly. Technical details point to Appendix H of arXiv 2511.16719.

Why it is moving now

The dated event is 27 March 2026. Teams already running SAM 3 on multi-object video (sports, traffic, annotation farms) hit linear cost as object count rises. Meta's release notes put a single-GPU number on that: about 7x speedup at 128 objects on one H100 versus the November 2025 SAM 3 release, plus listed inference optimizations (reduced CPU-GPU sync, torch.compile fusion, batched postprocessing). The Meta blog's 16-to-32 frames-per-second claim for a medium object count on one H100 is the same event's marketing page. Use the release-note tables for numbered benches.

License + activity + maturity (demo | usable | production-shaped)

License: the repository LICENSE file is titled "SAM License," last updated 19 November 2025. It grants a non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, copy, create derivative works of, and modify SAM Materials, with redistribution only under the same Agreement, a publication-acknowledgement duty, trade-control compliance, and a bar on reverse engineering underlying components. ITAR and specified military, nuclear, espionage, and illegal-weapons end uses are prohibited. Warranty is AS IS. California law and California courts. Litigation alleging IP infringement against Meta terminates the license.

Activity: tagged 27 March 2026 release notes and Hugging Face facebook/sam3.1. GitHub star count at publication is UNKNOWN. Weights are distributed through Hugging Face with authentication as described in the README getting-started section.

Maturity: usable for research and annotation pipelines that can accept SAM License, gated downloads, and an H100-class or similar GPU. Demo via Segment Anything Playground without local weights. Production-shaped only after counsel signs the SAM License, you pin a checkpoint digest, and you reproduce the 128-object timing on your hardware. Meta's "without sacrificing accuracy" is Meta's sentence.

When to pick it vs the default alternative

Pick SAM 3.1 when you already run SAM 3 video tracking, object counts are high, and you can pull the new checkpoints under SAM License.

Pick SAM 3 (November 2025) only if you cannot reinstall the repo. 3.1 needs current code.

Pick a closed video-analytics SaaS when you need a DPA and an SLA, not a research license.

Pick SAM 2 if you only need click-one-object tracking and cannot take SAM 3's concept prompts.

Don't pick it as Apache-2.0 weights for an unrestricted commercial binary. The LICENSE is not Apache.

Failure modes

Meta's 7x figure is at 128 objects on one H100 versus SAM 3. Your object count, GPU, and PyTorch version will differ. RELEASE_SAM3p1.md itself shows mixed SA-Co/VEval video PCS numbers (some public-benchmark columns down). VOS improved on 6 of 7 listed sets. YTVOS19 val went 89.7 to 89.3.

Gated Hugging Face access can block a deploy pipeline that expected anonymous wget. SAM License termination and California venue are real contract terms. Trade-control clauses bind redistribution. Outputs are AS IS. Medical or safety-critical overlay is your validation.

Links

Whether 7x reproduces on your GPU and object mix is UNKNOWN.