Netron

Visualizer for neural network, deep learning, and machine learning models

Free· Free and open-source (MIT license). Available as a web app, desktop app (macOS/Linux/Windows), and Python package at no cost.Evaluation

Visit website →

Best for

ML engineers, researchers, and students who need to inspect, debug, or teach the structure of neural network checkpoints across many frameworks.

Skip if

Anyone looking to train, fine-tune, edit, or benchmark models; Netron only visualizes existing model files, it does not modify or execute them.

Netron is an open-source viewer for neural network, deep learning, and machine learning models built by Lutz Roeder. Point it at a model file and it renders the computational graph as an interactive, zoomable diagram, letting you inspect every layer, operator, tensor shape, attribute, and weight without writing code. It runs entirely in the browser at netron.app (no upload required, models stay local), as a native desktop app for macOS/Linux/Windows, or as a Python package you can launch on a file or programmatically from a script or notebook.

Format coverage is unusually broad for a free tool. Stable support includes ONNX, TensorFlow Lite, PyTorch and TorchScript, torch.export and ExecuTorch, TensorFlow SavedModel/frozen graphs, Core ML, OpenVINO IR, Keras, Caffe, Darknet, Safetensors, and NumPy. Experimental parsers cover MLIR, JAX, GGUF (llama.cpp/quantized LLMs), RKNN, ncnn, MNN, PaddlePaddle, and scikit-learn, which makes Netron the de facto lingua franca when you need to peek inside a checkpoint from an unfamiliar framework.

Typical workflows: sanity-checking a converted ONNX export before deployment, debugging shape mismatches after quantization, comparing two versions of the same architecture side-by-side, reading a stranger's model.safetensors to figure out its tokenizer or embedding dims, teaching students what a transformer block actually looks like, or pulling a specific operator's attributes when writing a custom runtime kernel. It is a diagnostic and educational tool rather than a training or serving platform, but for that narrow job it is the tool almost every ML engineer already has in their bookmarks.

Editor's take

Netron is one of those quietly essential utilities that every ML practitioner ends up using. It is the fastest way to answer 'what is actually inside this .onnx / .pt / .gguf file?' without spinning up the right framework. Free, private (runs locally), and comprehensive, it earns a permanent place in the bookmark bar.

— The AI Tool Bible editorial team

Pros

✅ Broadest model-format coverage of any free viewer, from ONNX and PyTorch to GGUF and Safetensors
✅ Runs 100% client-side in the browser; model files never leave your machine
✅ Available as web app, native desktop app, and pip-installable Python package
✅ MIT-licensed, actively maintained by Lutz Roeder with frequent format updates
✅ Clean, interactive graph rendering with per-node attributes, tensor shapes, and weight previews
✅ Zero configuration: drag a file in and it just works, no build or conversion step
✅ Handles very large models (multi-GB LLM checkpoints) without crashing the browser tab

Cons

⚠️ Read-only viewer, no graph editing, patching, or model surgery
⚠️ Very large models can be slow to lay out and hard to navigate without collapsing subgraphs
⚠️ No built-in profiling, latency, or FLOP estimation, only structural inspection
⚠️ Experimental format parsers (JAX, MLIR, MNN, sklearn) can miss operators or fail on newer files
⚠️ Custom or vendor-specific ops render as opaque boxes with limited metadata
⚠️ No comparison/diff mode between two versions of the same model

Use cases

ONNX model inspectionPyTorch checkpoint debuggingQuantization shape verificationGGUF LLM structure inspectionSafetensors weight browsingModel architecture teachingCross-framework model conversion QACustom operator attribute lookupTensorFlow Lite mobile model reviewCore ML export validation