netron

Developed by Lutz Roeder and widely used in the ML research and engineering community, Netron renders the computational graph of a model showing each layer, its type, input/output dimensions, kernel sizes, and how data flows through the network in an interactive visual interface.

This makes it an essential tool for ML engineers and data scientists who need to inspect, debug, or document model architectures without writing custom visualization code.

Netron supports an exceptionally broad range of model formats: ONNX, TensorFlow Lite, PyTorch, TorchScript, ExecuTorch, torch.export, TensorFlow (SavedModel, Keras), Core ML (Apple), OpenVINO (Intel), Caffe, Darknet, Safetensors, NumPy, and more.

Experimental support extends to MLIR, JAX, GGUF (the format used by llama.cpp for quantized LLM deployment), RKNN, ncnn, MNN, PaddlePaddle, and scikit-learn.

This breadth of format support means Netron works across virtually every major ML framework and deployment target without requiring users to convert their model to a specific format first. Models can be exported as images for documentation, sharing, or publication purposes.

Netron is available as a web application at netron.app, as a desktop app for macOS, Windows, and Linux, and as a Python package installable via pip (pip install netron) for use within Jupyter notebooks and Python development environments.

It is free and open-source under a permissive license, with the source on GitHub.