The universal x-ray for neural network black boxes

What it does

Netron renders neural network models as interactive graphs. You drop in a file—ONNX, PyTorch, TensorFlow Lite, Keras, Core ML, even NumPy .npy files—and it draws the computation graph with nodes, edges, and metadata. It runs in a browser, as a desktop app, or via Python, and it works offline.

The interesting bit

The breadth is the trick. Netron doesn’t require you to install the framework that trained the model. It parses the serialized format directly, which means you can inspect a TensorFlow model without a working tf environment, or peek inside a Core ML model on Linux. The README lists experimental support for MLIR, JAX, GGUF, and scikit-learn—suggesting the author is actively chasing whatever format appears next.

Key highlights

• Supports 15+ formats including ONNX, PyTorch (multiple export paths), TensorFlow, Keras, Caffe, Darknet, Safetensors, OpenVINO, and NumPy
• Browser version loads models from URL parameters, making it trivial to share inspectable links
• Desktop builds for macOS, Linux, Windows; Python package for scripting workflows
• Experimental parsers for emerging formats like MLIR, ExecuTorch, and GGUF (the llama.cpp weight format)

Caveats

• “Experimental support” for roughly a third of the listed formats; depth of parsing may vary
• README doesn’t specify whether the viewer shows weights, just architecture, or execution metadata—clarity would help

Verdict

Essential if you inherit mystery models, debug conversion pipelines, or write framework tutorials. Skip it if you only train inside Jupyter and never look under the hood.