SAM 3: Open Segmentation Model (848M)

SAM 3 (Segment Anything Model 3) is Meta’s promptable segmentation model for images and videos. Official hosts we opened on 2026-10-13: gated weights at facebook/sam3, code at facebookresearch/sam3, and the project page ai.meta.com/sam3.

This page is a spoke from remove.bg alternatives. People search that commercial name when they want local, open cutouts. SAM 3 is broader: open-vocabulary detect, segment, and track with text phrases or visual prompts. It is not a hosted background-removal API. For pixel editors rather than mask models, see Photoshop alternatives.

The GitHub README (Meta Superintelligence Labs) says SAM 3 “can detect, segment, and track objects using text or visual prompts such as points, boxes, and masks.” Compared with SAM 2, it adds exhaustive segmentation of an open-vocabulary concept specified by a short text phrase or exemplars. The same README cites 75–80% of human performance on the SA-CO benchmark (270K unique concepts). That figure is upstream self-report; forkalt has not reproduced it.

What we verified (2026-10-13)

Field	Value	Source
Hugging Face likes	4,225	HF API facebook/sam3
HF downloads (monthly counter)	2,037,775	HF API
Card created	2025-11-07	HF createdAt
Card last modified	2025-11-20	HF lastModified
Gated	manual	HF API
Pipeline tag	mask-generation	HF API
License label	other	HF cardData
LICENSE file	SAM License, last updated 19 November 2025	GitHub
Parameters	848M	GitHub README “Model” section
GitHub stars / forks	11,952 / 1,821	GitHub API
GitHub created	2025-07-17	GitHub API
Homepage	https://ai.meta.com/sam3/	GitHub API

The Hugging Face README is gated (HTTP 401). Architecture and install facts are taken from the public GitHub README and LICENSE, plus public HF API fields.

How to run (from the official README)

Prerequisites listed in the README: Python 3.12+, PyTorch 2.7+, CUDA GPU with CUDA 12.6+. Install path: conda env, pip install torch==2.10.0 torchvision from the cu128 index, clone facebookresearch/sam3, pip install -e ..

⚠️ Checkpoints are gated. The README says to request access on facebook/sam3 and authenticate (hf auth login) before download.

Image path in the README: build_sam3_image_model() plus Sam3Processor, then set_text_prompt. Video path: build_sam3_video_predictor() with start_session / add_prompt requests. Example notebooks live under examples/ (image predictor, video predictor, batched inference, SAM 3 Agent).

Optional faster inference extras named in the README include flash-attn-3 and cc_torch. We have not timed them.

SAM 3.1 note

A “Latest updates” block dated 03/27/2026 says SAM 3.1 Object Multiplex checkpoints are on facebook/sam3.1. This URL stays SAM 3 (facebook/sam3). Do not treat SAM 3.1 likes or license as copied here.

License

Hugging Face license: other. GitHub license.spdx_id is NOASSERTION with name “Other”. The file is SAM License, last updated November 19, 2025, granting a non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, copy, create derivative works of, and modify the SAM Materials, subject to redistribution, research-acknowledgement, trade-control, and no-reverse-engineering clauses. Read the file. Not legal advice.

What we did not invent

• Peak VRAM and FPS on a named GPU are not yet verified.
• We do not claim SAM 3 matches remove.bg, Photoshop Select Subject, or SAM 2 on a particular benchmark unless the README states that comparison (it discusses SAM 2 capabilities, not a remove.bg test).
• We have not opened the gated HF README, so card-only claims beyond the public API are omitted.

The GitHub description field (API) is: “The repository provides code for running inference and finetuning with the Meta Segment Anything Model 3 (SAM 3), links for downloading the trained model checkpoints, and example notebooks that show how to use the model.” Last push observed in the 2026-10-13 GitHub check: 2026-10-07.

Architecture (official README)

“SAM 3 consists of a detector and a tracker that share a vision encoder. It has 848M parameters.” The detector is described as DETR-based and conditioned on text, geometry, and image. A presence token is called out as improving discrimination between close phrases (the README’s example: “a player in white” vs “a player in red”). A decoupled detector–tracker design is said to minimize task interference. Training data is described as an engine that annotated over 4 million unique concepts. The SA-CO image/video evaluation sets are hosted on Hugging Face as facebook/SACo-Gold, SACo-Silver, and SACo-VEval. forkalt has not downloaded those datasets.

The project page ai.meta.com/sam3 repeats text prompts, exemplar (box) prompts, visual click prompts, and interactive follow-ups. Instagram Edits / Meta AI Vibes mentions on that page are product roadmap copy from Meta, not an open-source shipping status we independently checked.

Examples in the repo

Notebooks listed in the README: sam3_image_predictor_example.ipynb, sam3_video_predictor_example.ipynb, sam3_image_batched_inference.ipynb, sam3_agent.ipynb, plus SA-CO visualization notebooks. Optional extras: pip install -e ".[notebooks]" and .[train,dev]. Paper, demo, and blog links are on the README header (ai.meta.com/research/publications/sam-3-segment-anything-with-concepts/, segment-anything.com, Meta blog).

See also Image & Design, remove.bg alternatives, Photoshop alternatives. Generation (not cutouts) is a different job: FLUX 3 Image Klein vs Qwen-Image-2.1-Turbo on Midjourney-class alternatives. Catalog: models.