YOLO26 ONNX

ONNX exports of the Ultralytics/YOLO26 checkpoints, one file per model size (n/s/m/l/x) and task (detection, instance segmentation, pose estimation, oriented bounding box, image classification).

All 25 files were exported with the Ultralytics exporter at imgsz=(480, 640), opset=12, FP32 with static input shapes, using the exporter defaults. The same 25 models are also shipped with fully dynamic batch, height and width axes under dynamic/, for batched/multi-camera consumers, and as compiled Qualcomm QNN context-binary models under qnn/, targeting the Hexagon NPU (HTP v75). Outputs follow the standard Ultralytics raw layouts listed below (classification outputs are already softmaxed probabilities). Class names are embedded in the ONNX graph metadata. The weights are unchanged from the earlier 640x640 exports - only the input geometry differs.

These files are derived from the Ultralytics YOLO26 checkpoints and inherit their license: AGPL-3.0. See https://www.ultralytics.com/license for the terms, and https://huggingface.co/Ultralytics/YOLO26 for the upstream weights.

Files

File Task Size Input Output Classes File size
yolo26n-cls.onnx image classification n images 1×3×480×640 output0 1×1000 float 1000 11.3 MB
yolo26s-cls.onnx image classification s images 1×3×480×640 output0 1×1000 float 1000 26.9 MB
yolo26m-cls.onnx image classification m images 1×3×480×640 output0 1×1000 float 1000 46.6 MB
yolo26l-cls.onnx image classification l images 1×3×480×640 output0 1×1000 float 1000 56.5 MB
yolo26x-cls.onnx image classification x images 1×3×480×640 output0 1×1000 float 1000 118.6 MB
yolo26n.onnx detection n images 1×3×480×640 output0 1×84×6300 float 80 9.9 MB
yolo26s.onnx detection s images 1×3×480×640 output0 1×84×6300 float 80 38.2 MB
yolo26m.onnx detection m images 1×3×480×640 output0 1×84×6300 float 80 81.9 MB
yolo26l.onnx detection l images 1×3×480×640 output0 1×84×6300 float 80 99.6 MB
yolo26x.onnx detection x images 1×3×480×640 output0 1×84×6300 float 80 223.2 MB
yolo26n-obb.onnx oriented bounding box n images 1×3×480×640 output0 1×20×6300 float 15 10.0 MB
yolo26s-obb.onnx oriented bounding box s images 1×3×480×640 output0 1×20×6300 float 15 39.2 MB
yolo26m-obb.onnx oriented bounding box m images 1×3×480×640 output0 1×20×6300 float 15 85.1 MB
yolo26l-obb.onnx oriented bounding box l images 1×3×480×640 output0 1×20×6300 float 15 102.7 MB
yolo26x-obb.onnx oriented bounding box x images 1×3×480×640 output0 1×20×6300 float 15 230.5 MB
yolo26n-pose.onnx pose estimation n images 1×3×480×640 output0 1×56×6300 float 1 12.0 MB
yolo26s-pose.onnx pose estimation s images 1×3×480×640 output0 1×56×6300 float 1 41.8 MB
yolo26m-pose.onnx pose estimation m images 1×3×480×640 output0 1×56×6300 float 1 86.5 MB
yolo26l-pose.onnx pose estimation l images 1×3×480×640 output0 1×56×6300 float 1 104.2 MB
yolo26x-pose.onnx pose estimation x images 1×3×480×640 output0 1×56×6300 float 1 230.6 MB
yolo26n-seg.onnx instance segmentation n images 1×3×480×640 output0 1×116×6300 float
output1 1×32×120×160 float
80 11.2 MB
yolo26s-seg.onnx instance segmentation s images 1×3×480×640 output0 1×116×6300 float
output1 1×32×120×160 float
80 41.9 MB
yolo26m-seg.onnx instance segmentation m images 1×3×480×640 output0 1×116×6300 float
output1 1×32×120×160 float
80 94.6 MB
yolo26l-seg.onnx instance segmentation l images 1×3×480×640 output0 1×116×6300 float
output1 1×32×120×160 float
80 112.2 MB
yolo26x-seg.onnx instance segmentation x images 1×3×480×640 output0 1×116×6300 float
output1 1×32×120×160 float
80 251.6 MB

Dynamic exports

The same 25 models also ship under dynamic/ with the batch, height and width axes all dynamic, for batched consumers (e.g. the multi-camera yolo_batch_node in mgonz13/yolo_ros). They are the same weights re-exported with dynamic=True: the input is ['batch', 3, 'height', 'width'], so any batch size and any input resolution that is a multiple of 32 is accepted. Class names are embedded in the metadata as before.

File Task Size Input Output Classes File size
dynamic/yolo26n-cls.onnx image classification n images batch×3×height×width output0 batch×1000 float 1000 11.3 MB
dynamic/yolo26s-cls.onnx image classification s images batch×3×height×width output0 batch×1000 float 1000 27.0 MB
dynamic/yolo26m-cls.onnx image classification m images batch×3×height×width output0 batch×1000 float 1000 46.6 MB
dynamic/yolo26l-cls.onnx image classification l images batch×3×height×width output0 batch×1000 float 1000 56.6 MB
dynamic/yolo26x-cls.onnx image classification x images batch×3×height×width output0 batch×1000 float 1000 118.7 MB
dynamic/yolo26n.onnx detection n images batch×3×height×width output0 batch×84×anchors float 80 10.4 MB
dynamic/yolo26s.onnx detection s images batch×3×height×width output0 batch×84×anchors float 80 38.6 MB
dynamic/yolo26m.onnx detection m images batch×3×height×width output0 batch×84×anchors float 80 82.3 MB
dynamic/yolo26l.onnx detection l images batch×3×height×width output0 batch×84×anchors float 80 100.8 MB
dynamic/yolo26x.onnx detection x images batch×3×height×width output0 batch×84×anchors float 80 223.6 MB
dynamic/yolo26n-obb.onnx oriented bounding box n images batch×3×height×width output0 batch×20×anchors float 15 10.9 MB
dynamic/yolo26s-obb.onnx oriented bounding box s images batch×3×height×width output0 batch×20×anchors float 15 39.6 MB
dynamic/yolo26m-obb.onnx oriented bounding box m images batch×3×height×width output0 batch×20×anchors float 15 85.6 MB
dynamic/yolo26l-obb.onnx oriented bounding box l images batch×3×height×width output0 batch×20×anchors float 15 103.2 MB
dynamic/yolo26x-obb.onnx oriented bounding box x images batch×3×height×width output0 batch×20×anchors float 15 232.3 MB
dynamic/yolo26n-pose.onnx pose estimation n images batch×3×height×width output0 batch×56×anchors float 1 12.1 MB
dynamic/yolo26s-pose.onnx pose estimation s images batch×3×height×width output0 batch×56×anchors float 1 41.9 MB
dynamic/yolo26m-pose.onnx pose estimation m images batch×3×height×width output0 batch×56×anchors float 1 86.7 MB
dynamic/yolo26l-pose.onnx pose estimation l images batch×3×height×width output0 batch×56×anchors float 1 104.5 MB
dynamic/yolo26x-pose.onnx pose estimation x images batch×3×height×width output0 batch×56×anchors float 1 231.3 MB
dynamic/yolo26n-seg.onnx instance segmentation n images batch×3×height×width output0 batch×116×anchors float
output1 batch×32×mask_height×mask_width float
80 11.5 MB
dynamic/yolo26s-seg.onnx instance segmentation s images batch×3×height×width output0 batch×116×anchors float
output1 batch×32×mask_height×mask_width float
80 42.2 MB
dynamic/yolo26m-seg.onnx instance segmentation m images batch×3×height×width output0 batch×116×anchors float
output1 batch×32×mask_height×mask_width float
80 95.0 MB
dynamic/yolo26l-seg.onnx instance segmentation l images batch×3×height×width output0 batch×116×anchors float
output1 batch×32×mask_height×mask_width float
80 114.1 MB
dynamic/yolo26x-seg.onnx instance segmentation x images batch×3×height×width output0 batch×116×anchors float
output1 batch×32×mask_height×mask_width float
80 252.1 MB

QNN exports

The same 25 checkpoints are also available as Qualcomm QNN models under qnn/. Each file is an ONNX Runtime EPContext model with the compiled QNN context binary embedded (no external context files), produced with the Ultralytics format=qnn exporter. The graph is quantized to W8A16 (8-bit weights, 16-bit activations) through ONNX Runtime's QNN QDQ flow, calibrated on the Ultralytics sample datasets, and finalized for the Hexagon Tensor Processor (HTP) v75 target (Snapdragon 8 Gen 3-class NPUs; the embedded context reports soc_type=QCS8300).

Inputs are static and channel-last (NHWC): images 1×640×480×3, and 1×640×640×3 for classification - a different geometry from the channel-first 1×3×480×640 FP32 exports above. Outputs follow the same raw layouts as the FP32 exports. These models only run through ONNX Runtime's QNN Execution Provider on Snapdragon hardware; the QNN context cannot run on the CPU or CUDA providers.

File Task Size Input Output Classes File size
qnn/yolo26n-cls_qnn.onnx image classification n images 1×640×640×3 output0 1×1000 float 1000 3.3 MB
qnn/yolo26s-cls_qnn.onnx image classification s images 1×640×640×3 output0 1×1000 float 1000 7.4 MB
qnn/yolo26m-cls_qnn.onnx image classification m images 1×640×640×3 output0 1×1000 float 1000 12.7 MB
qnn/yolo26l-cls_qnn.onnx image classification l images 1×640×640×3 output0 1×1000 float 1000 15.5 MB
qnn/yolo26x-cls_qnn.onnx image classification x images 1×640×640×3 output0 1×1000 float 1000 31.9 MB
qnn/yolo26n_qnn.onnx detection n images 1×640×480×3 output0 1×84×6300 float 80 3.4 MB
qnn/yolo26s_qnn.onnx detection s images 1×640×480×3 output0 1×84×6300 float 80 10.7 MB
qnn/yolo26m_qnn.onnx detection m images 1×640×480×3 output0 1×84×6300 float 80 22.1 MB
qnn/yolo26l_qnn.onnx detection l images 1×640×480×3 output0 1×84×6300 float 80 27.0 MB
qnn/yolo26x_qnn.onnx detection x images 1×640×480×3 output0 1×84×6300 float 80 59.2 MB
qnn/yolo26n-obb_qnn.onnx oriented bounding box n images 1×640×480×3 output0 1×20×6300 float 15 3.5 MB
qnn/yolo26s-obb_qnn.onnx oriented bounding box s images 1×640×480×3 output0 1×20×6300 float 15 11.0 MB
qnn/yolo26m-obb_qnn.onnx oriented bounding box m images 1×640×480×3 output0 1×20×6300 float 15 22.9 MB
qnn/yolo26l-obb_qnn.onnx oriented bounding box l images 1×640×480×3 output0 1×20×6300 float 15 27.8 MB
qnn/yolo26x-obb_qnn.onnx oriented bounding box x images 1×640×480×3 output0 1×20×6300 float 15 61.0 MB
qnn/yolo26n-pose_qnn.onnx pose estimation n images 1×640×480×3 output0 1×56×6300 float 1 4.0 MB
qnn/yolo26s-pose_qnn.onnx pose estimation s images 1×640×480×3 output0 1×56×6300 float 1 11.8 MB
qnn/yolo26m-pose_qnn.onnx pose estimation m images 1×640×480×3 output0 1×56×6300 float 1 23.5 MB
qnn/yolo26l-pose_qnn.onnx pose estimation l images 1×640×480×3 output0 1×56×6300 float 1 28.4 MB
qnn/yolo26x-pose_qnn.onnx pose estimation x images 1×640×480×3 output0 1×56×6300 float 1 61.1 MB
qnn/yolo26n-seg_qnn.onnx instance segmentation n images 1×640×480×3 output0 1×116×6300 float
output1 1×32×160×120 float
80 3.8 MB
qnn/yolo26s-seg_qnn.onnx instance segmentation s images 1×640×480×3 output0 1×116×6300 float
output1 1×32×160×120 float
80 11.7 MB
qnn/yolo26m-seg_qnn.onnx instance segmentation m images 1×640×480×3 output0 1×116×6300 float
output1 1×32×160×120 float
80 25.5 MB
qnn/yolo26l-seg_qnn.onnx instance segmentation l images 1×640×480×3 output0 1×116×6300 float
output1 1×32×160×120 float
80 30.4 MB
qnn/yolo26x-seg_qnn.onnx instance segmentation x images 1×640×480×3 output0 1×116×6300 float
output1 1×32×160×120 float
80 66.6 MB

Export recipe

uv run --with ultralytics --with onnx --with onnxruntime --with onnxslim \
  yolo export model=yolo26m.pt format=onnx imgsz=480,640 opset=12

Add dynamic=True for the dynamic/ variant: batch, height and width all stay dynamic, which is what the dynamic/ files ship (no post-export pinning step).

For the qnn/ variant use format=qnn name=75 instead - the model is quantized to W8A16 and compiled to an embedded QNN context binary for the HTP v75 target (no Snapdragon device is needed for the export step):

uv run --with ultralytics --with onnxruntime-qnn \
  yolo export model=yolo26m.pt format=qnn name=75 imgsz=640,480

Use imgsz=640 for classification. onnxruntime-qnn bundles the Qualcomm AI Runtime (QAIRT) libraries and is installed automatically on first QNN export.

The .pt checkpoint names follow yolo26{size}{-task} (no task suffix for detection; -seg, -pose, -obb, -cls for the other tasks); the exported .onnx keeps the same stem. Replace yolo26m.pt with the checkpoint of interest.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for unileon-robotics/YOLO26-ONNX

Quantized
(60)
this model

Collection including unileon-robotics/YOLO26-ONNX