// the find
lutzroeder/netron
Visualizer for neural network, deep learning and machine learning models
Netron is a model viewer that opens a neural network file and renders it as an interactive graph of layers, weights, and shapes. It's for anyone who needs to inspect a trained model's architecture without writing code — ML engineers debugging export issues, reverse engineers poking at someone else's .onnx or .gguf file, or anyone who just wants to see what's inside a checkpoint.
Format coverage is absurd in a good way — ONNX, TensorFlow Lite, PyTorch, Core ML, GGUF, safetensors, and two dozen more, each with its own metadata/schema file, so it's not a thin wrapper around one protobuf parser. Ships as a browser app, pip package, and native installer for macOS/Windows/Linux, so it fits whatever workflow you're already in instead of forcing Electron on you. Runs entirely client-side — no upload, no server round-trip — which matters when the model file is something you can't legally send to a third party.
It's a viewer, not an editor — you can't modify weights, prune nodes, or export a changed graph, so it's a dead end if inspection turns into actual surgery. Format support is maintained per-parser by one person (lutzroeder), so a new or obscure format can lag for a while, and the 'experimental' tag on MLIR/JAX/RKNN/ncnn/MNN is a real caveat, not boilerplate. Large models (anything multi-GB) get slow and memory-heavy in the graph renderer since it's laying out the whole thing as DOM/SVG, not a virtualized canvas.