Drop the model
Any size; the parser reads the protobuf directly and does not load weights into memory beyond their sizes.
Models
ONNX Model Inspector
The first question about any exported model is what it expects and what it returns. Drop an .onnx file and the inspector parses it in the browser: input and output names with shapes and types, dynamic dimensions, the operator histogram, the largest weights, opset and producer, and metadata such as class names that exporters embed.
Loading the tool…
How it works
Read the summary
Inputs and outputs first, then ops by count, then initializers by size, then metadata.
Export
A JSON summary, or a Markdown snippet for a model card.
Formats, limits and the numbers that matter
Dynamic axes
Dimensions with a name instead of a number (batch, height) are dynamic; runtimes need a concrete value at inference. The inspector lists them.
YOLO exports
Ultralytics ONNX files carry names, stride and imgsz in metadata_props; the inspector shows them so you can set up postprocessing.
External data
Models over 2 GB store weights in a side file; the inspector reads the graph and reports the external references.
Common questions
Can it run the model?
Not in this version; running with onnxruntime-web on a sample input is planned.
TensorFlow or PyTorch files?
No; export to ONNX first, which both frameworks support.
Is the file uploaded?
No; parsing happens in the page.