ultralytics-yolo26

Pretrained Ultralytics YOLO26 detection weights, converted from the original .pt checkpoints to safetensors, for use with vision-rs.

Contents

ultralytics/yolo26/
├── yolo26n.safetensors
├── yolo26s.safetensors
├── yolo26m.safetensors
├── yolo26l.safetensors
└── yolo26x.safetensors

Variant	Params	Source file (Ultralytics release v8.4.0)
n	2,591,962	https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n.pt
s	10,047,882	https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26s.pt
m	21,952,564	https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26m.pt
l	26,370,410	https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26l.pt
x	59,099,338	https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26x.pt

Conversion

Each .pt checkpoint was loaded with torch.load(..., weights_only=False), its state_dict extracted, every tensor cast to float32 and made contiguous, then written out with safetensors.torch.save_file. Tensor keys are the original Ultralytics module names (e.g. model.0.conv.weight) — they have not been remapped to vision-rs's internal parameter names. See scripts/ultralytics/convert_model.py in the vision-rs repo for the exact conversion logic.

License

These weights are derived from Ultralytics YOLO26, which is released under AGPL-3.0. See the LICENSE file in this repository. Ultralytics also offers a commercial/Enterprise license for users who cannot comply with AGPL-3.0 — see https://www.ultralytics.com/license for details.

Usage

These files are consumed automatically by vision-rs's model download pipeline (assets/models/ultralytics/yolo26{n,s,m,l,x}.toml); you generally shouldn't need to download them by hand. To fetch one directly:

huggingface-cli download teenygrad/ultralytics-yolo26 \
  ultralytics/yolo26/yolo26n.safetensors --local-dir .