The ONNX Model Zoo stopped serving weights on 1 July 2025 and points at Hugging Face instead

LFS downloads ended on 1 July 2025 and the notice comes first

The first thing the README says is a deprecation notice. It thanks the community for the Model Zoo effort, explains that much of the novel model sharing has transitioned to Hugging Face, and states that the repository is preserved for historical purposes only. The hard detail is a date: models will no longer be available for LFS download starting 1 July 2025. Models that were originally available here are still reachable at:

https://huggingface.co/onnxmodelzoo

The consequence for anyone who clones this repository today is predictable and quiet. Model files are held in Git LFS because they can be large, and with LFS downloads retired you should not expect to fetch the weights from this repository. Fetching the weights from the Hugging Face space first, or after, is the step the notice is really asking for.

Two tiers of trust sit in one tree, validated/ and four topical folders

The repository root mixes two different things. There are topical directories, Computer_Vision/, Natural_Language_Processing/, Generative_AI/, and Graph_Machine_Learning/, which are the categories being expanded as new models arrive. There is also a validated/ directory holding the models that went through accuracy validation, and the README is explicit that new models should be judged against that subset rather than the whole collection, because the incoming ones are still being rigorously checked.

The consequence is that the directory name tells you the provenance of a model, not its quality. A model sitting directly under Computer_Vision/ is not necessarily wrong, it is simply unvalidated, and there is no marker in a path that distinguishes one from the other at a glance. For a team that has to answer what is production-ready, this repository supplies the validated subset as the answer.

The accuracy column reports paper numbers, not measurements from this tree

The image classification table carries a Top-5 error figure for each model, and every one is attributed to the paper. MobileNet is listed at about 10 percent, ResNet at about 3.6 percent, SqueezeNet at about 20 percent, VGG at about 8 percent, AlexNet at about 15 percent, and GoogleNet at about 6.7 percent. Layer counts appear in the same descriptions, from AlexNet at up to 8 layers through VGG at up to 19, GoogleNet at up to 22, and ResNet at up to 152.

The consequence is that these are references, not a benchmark table. They describe what the authors reported on their own setup, they say nothing about how the exported ONNX file behaves after quantization, and they carry no hardware or runtime context. What the project offers instead is a validation path: starter Python code that checks your ONNX model against test data, which is the only way these numbers become yours.

INT8 files come from a separate Intel toolchain, not from a flag here

The README attributes every INT8 model to Intel Neural Compressor, an open source Python library whose selling point is accuracy driven tuning strategies that find a good quantized model quickly. It implements dynamic and static quantization for ONNX models and can represent them in operator oriented form or as tensor oriented QDQ pairs. You can drive it from a web based UI service or from Python code.

That separation has a practical consequence. Nothing in this repository performs quantization, so a quantized variant is a second artifact produced by a second toolchain against the same source model, and the two artifacts have to be kept in step yourself. If you validated the floating point export and then swap in the INT8 one, you are making an accuracy decision that the paper numbers do not cover. Validate the quantized file with the same starter code, on the same test data, or you are guessing.

Three file extensions share one repository and different runtimes

The Usage section notes three file formats inside the Model Zoo: .onnx, .pb, and .npz. The README points to that section for details on each format.

The consequence is that the format, not the category, decides what a downstream runtime can open. Model paths follow the same logic, and the classification collection nests some families in subdirectories rather than sitting flat:

validated/vision/classification/mobilenet
validated/vision/classification/inception_and_googlenet/googlenet

Inception and GoogLeNet share a parent folder, so a script that globs one level deep under classification will miss them.

Exports trace back to timm, torchvision, torch hub, and transformers

The models are not trained here. They are sourced from prominent open source repositories, with timm, torchvision, torch hub, and transformers named, and exported into the ONNX format using the open source TurnkeyML toolchain. That provenance has two consequences. First, parity with the upstream weights and preprocessing configuration is inherited, including the subtleties that come with it.

The README lists Computer Vision, NLP, Generative AI and Graph Machine Learning together as categories the project is currently expanding.

No releases and no tags means the only pin is a commit

This repository has no GitHub releases, so there is no version string to depend on. What you can anchor to is a commit, with the last push landing on 1 November 2026. Combined with the deprecation notice, that shapes how to use it: the tree is frozen in the only sense that matters for planning, since new model sharing moved to Hugging Face and accuracy validation is not the activity this repository is being kept for.

The consequence is a two part answer to where weights come from. For provenance, a model path under validated/ plus the reference in the table points you to the paper the model comes from. For the file itself, you fetch from the onnxmodelzoo space. Anyone who writes a build step against this tree should record the commit hash alongside the model name, because a path that resolves today tells a reader nothing about what they will get next month.

Editorial conclusion

Use this repository as documentation and as a record of which models were validated for accuracy, then fetch weights from the onnxmodelzoo space on Hugging Face. It does not suit anyone expecting new model additions or a released version to pin, because the tree is preserved rather than extended and the repository has no releases. Before building a pipeline on it, check that your runtime accepts the .onnx, .pb, or .npz file you need, and remember the accuracy figures in the tables are the values reported in the source papers.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. onnx/models on GitHub
4. README