Computer vision models and pretrained weights, free to download

How to load these weights

pip install ultralytics

from ultralytics import YOLO

# Downloads on first use, or pass a local path to your .pt file
model = YOLO("yolo11n.pt")
results = model("image.jpg")

Every model family

Jump straight to any family for its full write-up, FAQ and download table.

• YOLO26 weights
• YOLO11 weights
• YOLO12 weights
• YOLOv10 weights
• YOLOv9 weights
• YOLOv8 weights
• YOLOv5 (v5u) weights
• RT-DETR weights
• YOLO-World weights
• YOLOE weights
• YOLO26-seg weights
• YOLO26-pose weights
• YOLO26-obb weights
• YOLO26-cls weights
• SAM 2.1 weights
• SAM weights
• MobileSAM weights
• FastSAM weights
• Depth Anything V2 weights

Frequently asked questions

Are these computer-vision model weights free to download?

Yes. Every checkpoint here links to the official, publicly hosted weight file, with no sign-up. What differs is the license you may use them under. Most Ultralytics YOLO weights are AGPL-3.0 (free for open-source, needs an Enterprise license for closed commercial use), while SAM, SAM 2.1, MobileSAM and RT-DETR ship under Apache-2.0. Free to download is not the same as free to ship, though. Depth Anything V2's Base and Large checkpoints are CC-BY-NC-4.0, meaning non-commercial only, even though the download itself is public. And loading an Apache-2.0 model through the Ultralytics package brings that package's AGPL-3.0 terms with it. Check the license badge on the exact variant before shipping.

How do I download YOLO, SAM or RT-DETR weights here?

Find the model in the table, then click the Download button on the variant you want. We first verify the file is available and then start the download straight from the official CDN, showing a confirmation popup, or the exact error if something goes wrong. You can also copy the filename and let the Ultralytics library fetch it automatically on first use.

What is a .pt file and how do I load it?

A .pt file is a PyTorch checkpoint containing the trained model weights. For any Ultralytics-ecosystem model you can load it in two lines: `from ultralytics import YOLO` then `model = YOLO("yolo11n.pt")`. On first use the library also downloads the file automatically, so the direct links here are for offline installs, air-gapped machines and mirroring. Not everything here is an Ultralytics model, though: Depth Anything V2 ships .pth checkpoints that load through its own repo, and each family page shows the loader that actually works for it.

Which model weights should I download for my project?

For general object detection start with YOLO11n or YOLO26n: small, fast, well-supported. Need commercial-friendly licensing? Use RT-DETR or SAM 2.1 (Apache-2.0). Need to segment or track anything from a prompt? Use SAM 2.1, or MobileSAM and FastSAM for the edge. Detect classes by text with no training? Use YOLO-World. The nano (n) or tiny variant is almost always the right first download.

What accuracy do these numbers represent?

For the detectors, the score is COCO val2017 mAP50-95 at the input size listed, as published by each model's own authors, not a benchmark re-run on this site. Cross-family numbers are indicative of tier, not a controlled head-to-head, because the families use different size ladders and training recipes. The Segment Anything models are promptable, so we list parameters and download size instead of a single accuracy figure.

Do I need internet access to use these after downloading?

No. Once you have the .pt file locally you can point the loader at the file path and run fully offline, which is exactly why these direct links exist. Weights are self-contained; only the first download needs a connection.