SAM 3 Download Get Meta’s Segment Anything Model 3 (Official Weights & Code)

Download SAM 3 the latest breakthrough from Meta AI and unlock open-vocabulary segmentation for both images and video. With official weights, full code, and promptable concept tracking, you can go from idea to implementation in minutes. No limits. No labels. Just powerful vision AI yours to explore.

Start Creating Free Watch Demo
SAM 3 Download

SAM 3 Download: Complete Guide to Meta AI’s Latest Segmentation Model

Meta AI’s Segment Anything Model 3 (SAM 3) is one of the most advanced vision foundation models yet released offering open‑vocabulary segmentation, promptable concept detection, and instance tracking across images and video. Whether you’re a developer, researcher, engineer, or creator, downloading and using SAM 3 effectively unlocks powerful capabilities for computer vision tasks.

This guide covers everything you need to know about “SAM 3 download”  from official sources and checkpoints to installation, usage, licensing, integration options, advanced workflows, troubleshooting, and best practices.

 

1. What Is SAM 3?

SAM 3 (Segment Anything Model 3) is the latest segmentation foundation model from Meta AI. Unlike older segmentation models trained on fixed label sets, SAM 3 is open‑vocabulary  meaning it can interpret text prompts, image exemplars, or hybrid prompts (text + image) to segment and track all relevant instances of a concept within an image or video.

Key features include:

  • Text prompt segmentation (“red car”, “children with hats”)

  • Image exemplar segmentation (visual example as the query)

  • Hybrid prompting

  • Multi‑instance output

  • Video tracking with consistent IDs

  • No hardcoded class list

Downloading SAM 3 gives you access to both the model code and pretrained weights needed to run these capabilities locally or on a server.

2. Why SAM 3 Matters

Traditional models like Mask R‑CNN or DeepLab are limited by their class vocabularies. SAM 3’s openness means:

  • You can segment any concept, even unseen classes

  • It generalizes across domains (everyday objects, outdoor scenes, industrial parts)

  • Prompts are intuitive (e.g., text, example image)

  • It’s state‑of‑the‑art for segmentation and tracking workflows

For developers and businesses, the ability to download and run SAM 3 locally or on cloud GPUs means integrating powerful vision features into products without relying on cloud APIs.

3. Official Sources for SAM 3 Download

There are three primary official sources to download SAM 3:

3.1 Meta Research Release Page

Meta publishes research announcements on its site, often including:

  • Paper PDFs

  • Dataset/benchmark info

  • Download links for code and models

This is the canonical source to verify authenticity and license information.

3.2 GitHub Repository (facebookresearch/sam3)

The official repository hosted by Meta’s research organization typically includes:

  • Model source code

  • Example notebooks

  • Scripts for inference/training

  • Links to checkpoint downloads

This is the most developer‑friendly route for direct downloads.

📌 Search facebookresearch sam3 GitHub for the official repo.

3.3 Hugging Face Model Hub

SAM 3 is also published on Hugging Face under the model card:

  • Code snippets

  • Transformers integration

  • Hosted model weights (subject to access requirements)

  • Community examples & spaces

Model cards often link to weights and usage instructions.

4. Understanding Model Checkpoints

4.1 What Are Checkpoints?

A checkpoint is a file (or set of files) representing the saved pretrained weights of a model at a certain point. Downloading SAM 3 requires obtaining these checkpoints.

Checkpoints contain:

  • Learned parameters

  • Model configuration

  • Metadata

SAM 3 checkpoints can be large often multiple GB and may come in compressed formats (.pt, .bin, etc.).

4.2 How to Download Checkpoints

Checkpoints are typically hosted in one of the following:

  • Direct links on GitHub (via releases)

  • Cloud storage links (Google Drive, AWS S3)

  • Hugging Face Model Hub artifacts

Important: Some checkpoints may require agreeing to Meta’s terms or submitting an access request.

4.3 File Sizes and Storage Needs

Expect:

  • Primary model weights: Gigabytes in size

  • Auxiliary files: Smaller but necessary for code execution

  • Video tracking modules (if separate): Additional space

Always ensure you have enough SSD storage to host model files + temp files for inference.

5. Licensing and Access Requirements

SAM 3 is subject to Meta’s licensing rules. Depending on how and where you download:

  • You may need to agree to an end‑user license

  • Check for use restrictions (commercial vs research)

  • Hugging Face weights may require a login token

  • Always read the model card license before downloading

Downloading unauthorized replicas or weights from unverified sources can pose legal and security risks.

6. Setting Up Your Environment

Before running SAM 3 after download, prepare your system.

6.1 Python & Libraries

Common requirements:

 
pip install torch torchvision torchaudio pip install transformers pip install opencv python pillow

Other useful packages:

  • huggingface_hub

  • datasets (for custom training)

Match PyTorch and CUDA versions to your GPU for acceleration.

6.2 Hardware Considerations

SAM 3 inference and especially video tracking benefit from:

  • Powerful GPU (e.g., NVIDIA RTX class)

  • Ample VRAM (8–24+ GB)

  • CUDA & cuDNN installed

CPU‑only setups are possible but slow.

7. Downloading SAM 3 from GitHub

7.1 Step 1: Clone the Repository
 
git clone https://github.com/facebookresearch/sam3.git cd sam3

This pulls all source files, examples, and documentation.

7.2 Step 2: Install Dependencies

Most repos include a requirements.txt:

 
pip install -r requirements.txt

Or use a virtual environment:

 
python -m venv sam3-env source sam3-env/bin/activate pip install -r requirements.txt
7.3 Step 3: Download Model Weights

The repository linking page should include checkpoint URLs, such as:

  • sam3_base.pt

  • sam3_large.pt

Follow the README instructions to place weights in the expected directory (often ./weights/).

7.4 Step 4: Validate Your Setup

Many repos include a small sample image and a script:

 
python demo_image.py --image example.jpg --prompt "red car"

Ensure this runs without errors before proceeding.

8. Using SAM 3 via Hugging Face

8.1 Transformers Integration

Hugging Face supports SAM 3 in the transformers library:

 
from transformers import Sam3Processor, Sam3Model

This abstracts preprocessing and inference, making downloads easier.

Example:

 
processor = Sam3Processor.from_pretrained("facebook/sam3") model = Sam3Model.from_pretrained("facebook/sam3")

This auto‑downloads weights to your Hugging Face cache.

8.2 Hugging Face CLI Login
 
huggingface-cli login

Enter your API token so downloads are authorized.

8.3 Usage Example
 
from PIL import Image img = Image.open("scene.jpg") inputs = processor(images=img, text="blue backpack", return_tensors="pt") outputs = model(**inputs)

This returns segmentation outputs you can visualize.

9. Running Your First SAM 3 Inference

After downloading and installing:

9.1 Command‑Line

Some repos include CLI scripts:

 
python run_sam3.py --image test.jpg --prompt "person on bike"

Expected output: mask overlays or JSON data.

9.2 Python Workflow
 
from sam3 import Sam3Model model = Sam3Model.from_pretrained("facebook/sam3") img = load_image("img.jpg") result = model.segment_with_prompt(img, "white dog")

Visualize:

 
display_mask(img, result.masks[0])

10. Text Prompt Segmentation Downloads

SAM 3 excels with text prompts — enabling open‑vocabulary segmentation.

Examples:

  • “yellow school bus”

  • “people holding phones”

  • “blue umbrella”

Downloading SAM 3 weights lets you perform these tasks offline.

11. Video Tracking and Multi‑Frame Use

Promptable concept segmentation extends to video:

  1. Provide an initial prompt (text or image)

  2. SAM 3 identifies instances

  3. Model tracks instances across frames

Check the official demos and notebooks usually included in GitHub for video inference scripts.

12. Hybrid Prompt Workflows

Hybrid prompts combine:

  • Text (“green chair”)

  • Image exemplar (visual sample)

This improves precision for niche or ambiguous objects.

Downloaded SAM 3 supports these workflows if you use the appropriate API calls.

13. Fine‑Tuning and Domain‑Specific Downloads

While the base SAM 3 weights are general‑purpose, you can:

  • Fine‑tune on your own dataset

  • Generate custom checkpoints

  • Host them on Hugging Face

This is useful for:

  • Medical imaging

  • Industrial inspection

  • Satellite/aerial imagery

Fine‑tuning helps overcome domain gaps.

14. Integration with Tools & Platforms

14.1 Ultralytics

Ultralytics YOLO tools integrate SAM 3 for segmentation pipelines.

14.2 Annotation Tools

Platforms like CVAT or Label Studio can integrate SAM 3 outputs to automate labeling.

14.3 No‑Code Tools

Tools like ComfyUI support SAM 3 nodes for visual workflow automation.

15. Common Download Issues and Fixes

Issue: Download Fails / Auth Required

  • Ensure you agreed to terms

  • Check Hugging Face login

  • Verify token scopes

Issue: Checkpoint Too Large

Use tools like aria2c or direct download managers for stable download.

Issue: CUDA Errors

Match PyTorch to correct CUDA version.

16. Best Practices After Download

  • Cache models locally to avoid repeated downloads

  • Use hybrid prompts for ambiguity

  • Monitor GPU VRAM for large images/videos

  • Use prompt engineering to refine masks

17. Security & Ethics

Downloaded models can be used for:

  • Object tracking ensure privacy compliance

  • Sensitive data respect ethical boundaries

Always follow legal standards.

18. Benchmarking and SA‑Co

SA‑Co (Segment Anything with Concepts) is the official benchmark for evaluating open‑vocabulary segmentation and tracking. Use scripts in the GitHub repo to run evaluations.

19. Future Directions for SAM 3

Expect:

  • Real‑time edge deployments

  • Smaller lightweight variants

  • Conversational prompt interfaces

  • Open commercial APIs

20. Summary

Downloading SAM 3 unlocks:

  • Open‑vocabulary segmentation

  • Text/image/hybrid prompting

  • Video tracking

  • Integration with codebases & tools

Start by visiting the official GitHub or Hugging Face pages, download the checkpoints, set up your environment, and you’re ready to build powerful vision workflows.


AI RESEARCH FROM META

Introducing Segment Anything Model 3 (SAM 3) - the future of segmentation is promptable. Use text or visual prompts to instantly identify, segment, and track any object in images or video. Coming soon to Instagram Edits and Meta AI's Vibes.

Start Creating Free Download the model Try Playground