SAM 3.1
Release in the Segment Anything (SAM) family · version 3.1 (Object Multiplex)
Maintained by Meta (Meta Superintelligence Labs)34
SAM 3.1 is a March 2026 update of Meta's SAM 3, a model that detects, segments, and tracks objects in images and video from text phrases, image exemplars, or visual prompts such as points, boxes, and masks. The update adds Object Multiplex, which groups tracked objects into shared-memory buckets and processes them jointly instead of one at a time, together with new checkpoints and inference optimizations.431
- Model hub: Model checkpoints (Hugging Face, gated) (external site: huggingface.co)
- Repository: SAM 3 repository (external site: github.com)
- Release notes: SAM 3.1 release notes (external site: github.com)
- License: SAM License (external site: github.com)
- Paper: SAM 3: Segment Anything with Concepts (arXiv 2511.16719) (external site: arxiv.org)
Availability and license
Overall availability
Checkpoints are gated on Hugging Face: users must submit contact information and request access, and downloads work only after the request is accepted. The Hugging Face repository holds checkpoints only and has no Transformers integration; the code is in the SAM 3 GitHub repository.123
Availability is separate from permission: read the license before using or redistributing.
Custom Meta license (last updated November 19, 2025) covering code, weights, and documentation. It grants a non-exclusive, worldwide, non-transferable, royalty-free license to use, reproduce, modify, and distribute; redistributions must carry the agreement, and published research must acknowledge use of the SAM Materials. Users must comply with trade controls, may not be targets of sanctions, and may not use the materials for ITAR-regulated activities or prohibited end uses including military or warfare, nuclear, espionage, or weapons applications. It forbids reverse engineering, ends for anyone who brings IP litigation against Meta over the materials, is governed by California law, and may be modified by Meta. This differs from the Apache 2.0 license used for SAM 2.510
Model-disclosure tier
The weights can be obtained only by request, with approval, or by some users — for example a gated download that the publisher reviews. Not counted as open-weight.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Partial | Hugging Face access is approval-gated (manual review of access requests).23 |
| Inference codeIs code for running the model published? | Public | The SAM 3 repository provides image and video predictors and example notebooks; SAM 3.1 checkpoints require the latest repository code.3 |
| Training codeIs the code used to train the model published? | Unknown | The repository publishes only fine-tuning code (train.py with Hydra configs) for custom datasets, and Meta's announcement describes it as fine-tuning code. Under the rubric, fine-tuning code does not count as training code for the base model, and this review did not find the pretraining pipeline released.69 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The SAM 3 paper describes a training set with 4 million unique concept labels built by a data engine. The SA-Co evaluation benchmarks are released; this review did not find the training set itself released.83 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The paper and release notes describe the architecture and the Object Multiplex approach (paper Appendix H); no complete SAM 3.1 training configuration was located.84 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The release notes report SAM 3.1 results. The SA-Co benchmarks and evaluation code with configurations for reproducing SAM 3 results are published, but SAM 3.1-specific evaluation configurations were not located.47 |
What it is useful for
Run and use notes
- The README lists Python 3.12 or higher, PyTorch 2.7 or higher, and a CUDA-compatible GPU with CUDA 12.6 or higher as prerequisites, and requires authenticating to Hugging Face to download approved checkpoints.3
Organization context
Provenance and derivatives
An updated checkpoint of Meta's SAM 3 model with new multi-object tracking; SAM 3's tracker inherits the SAM 2 transformer encoder-decoder architecture.43
- Derived from: SAM 3 (November 2025 release) (external site: huggingface.co) — Predecessor checkpoint from the same developer; no separate catalog record.
Other releases in the Segment Anything (SAM) family
- SAM 2.1 Hiera-LargeModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: U.S. headquarters
Released by Meta (the SAM 3 repository credits Meta Superintelligence Labs) in the facebookresearch GitHub organization and the facebook Hugging Face account. Meta Platforms, Inc. has its principal executive offices in Menlo Park, California, per its Form 10-K. The SAM License names Meta Platforms Ireland Limited as licensor for EEA and Swiss users and Meta Platforms, Inc. for everyone else.3511
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.