Neural network model viewer and edge AI analyzer
Open ONNX, TFLite, PyTorch, GGUF and Safetensors models in your browser, then see whether they will run on your edge hardware. ModelVisio is free and open source (MIT). Parsing runs in a Web Worker in your browser, so the model file never leaves your machine.
Download the desktop app for macOS, Windows or Linux, use the VS Code extension, or view the source on GitHub.
View any model graph
A Netron-style interactive graph with a per-layer inspector showing shapes, weights, FLOPs/MACs and quantization sensitivity. Export the graph to SVG or PNG.
Will it compile? Op-support pre-flight for Coral, Hailo, RKNN and Kneron
Rule-based op-support checks against each vendor's documented limits (Coral Edge TPU op list and tensor rules; data-dependent ops and dynamic shapes on RKNN, Hailo and Kneron), plus memory warnings. Undoable auto-fixes include SiLU to HardSwish, SPPF to parallel SPP, Resize to ConvTranspose and channel pruning. The vendor compiler still has the final word. See which ops each NPU rejects.
Estimated FPS, FPS/W and memory fit on 21 edge accelerators
A 0 to 100 hardware-fit score from the roofline model (Williams, Waterman and Patterson, 2009): estimated FPS, FPS/W, whether the model is compute- or memory-bound, and a memory-fit check. Jetson AGX Orin and Hailo-8 are calibrated from MLPerf and vendor ResNet-50 numbers; other devices are estimates and are labelled as such. A model that does not fit in device memory scores at most 15. Compare the specs in the edge AI accelerator table.
Optimize and benchmark
Ranked, device-specific recommendations with estimated latency and size impact, a per-layer INT8/INT4 quantization heatmap, a Pareto search over device, runtime, precision and input size, and on-device ONNX benchmarks with onnxruntime-web (WebAssembly, WebGPU or WebNN).
Ship: conversion, deploy scripts and CI
Graph JSON, Layers CSV, Safetensors and NumPy exports, ready-to-run TensorRT, HailoRT and RKNN deployment scripts, a deployment agent that plans from a goal such as "Jetson Orin Nano at 30 FPS", fleet simulation, a local model registry, and a CLI and GitHub Action with budget and regression gates.
An AI copilot that reads your model's analysis
Ask Google Gemini or xAI Grok about your model. The copilot is given the computed analysis and cites live sources. The API key stays on the server and never reaches the browser.
Supported formats
Fully parsed: ONNX, TFLite, PyTorch (.pt/.pth), Safetensors, GGUF, NumPy, Darknet, Core ML, OpenVINO IR, TensorFlow, Keras, Caffe, PaddlePaddle, ncnn, MNN, RKNN, MLIR and scikit-learn. Detected and shown as metadata: TensorRT engines, TorchScript, ExecuTorch, compiled Core ML and more. All supported formats.
Frequently asked questions
- Is my model uploaded anywhere?
- No. The model file is parsed locally in your browser and never uploaded. The web app may send basic usage metadata (file name, format, size, parameter and layer counts) to its server for usage logging, and messages you send to the optional AI copilot go to its server-side proxy.
- How is ModelVisio different from Netron?
- Both draw the model graph. ModelVisio adds edge deployment analysis on top: compiler pre-flight, hardware-fit scoring for 21 accelerators, auto-fixes, benchmarks and deploy scripts.
- How accurate are the FPS estimates?
- They are roofline estimates. Calibrated devices use published MLPerf and vendor numbers; the rest are marked as estimated. Benchmark on the real device before you ship.
- What does it cost?
- Nothing. ModelVisio is MIT-licensed open source.
ModelVisio 1.2.0 · Developed by Premchand · Changelog · MIT License