ONNX Model Zoo: A Deprecated Archive of Pre-Trained Models

What the ONNX Model Zoo Was and Why It Mattered

The ONNX Model Zoo was a centralized repository that collected pre-trained, state-of-the-art machine learning models in the Open Neural Network Exchange format. ONNX is an open standard format created to represent machine learning models, defining a common set of operators and a common file format so that developers could move models between different frameworks, tools, runtimes, and compilers without rewriting them.

Before the ecosystem settled on Hugging Face as the primary model-sharing hub, the ONNX Model Zoo gave developers a single, vetted collection to draw from. Models were sourced from prominent open-source repositories including timm, torchvision, torch_hub, and transformers, and exported into the ONNX format using the TurnkeyML toolchain. This workflow let the team validate accuracy before publishing.

The repository served computer vision engineers who needed image classification, object detection, and body analysis models ready to drop into a runtime without a training pipeline, as well as NLP developers who needed machine comprehension and translation models in a format compatible with hardware accelerators. For a time it was the canonical answer to the question of where to find a validated ONNX ResNet or MobileNet.

Repository Layout and Model Categories

The repository is organized by model domain. The top-level directories are Computer_Vision, Generative_AI, Graph_Machine_Learning, and Natural_Language_Processing, mirroring the four categories the team was expanding when deprecation was announced.

Within those directories, the validated models are further organized by task. Under vision, the categories are image classification, object detection and image segmentation, body and face and gesture analysis, and image manipulation. Under language, the categories are machine comprehension, machine translation, and language modelling. The repository also contains validated models for visual question answering, speech and audio processing, and other tasks.

Image classification is the most populated category. It includes MobileNet, ResNet, SqueezeNet, VGG, AlexNet, GoogleNet, CaffeNet, DenseNet-121, and others, each linked to the original research paper. The validated directory contains the models that passed the team's accuracy checks. An ONNX_HUB_MANIFEST.json file at the root documents the collection in a machine-readable format.

Accessing Models After the Deprecation

The repository's deprecation notice states clearly that models are no longer available for LFS download starting July 1, 2025. The team has migrated the model collection to Hugging Face at https://huggingface.co/onnxmodelzoo.

The models that were stored in the repository using Git LFS are gone from the GitHub-hosted storage. The repository files and directory structure remain accessible for browsing, but the large binary model files (.onnx, .pb) are no longer downloadable from GitHub. Developers who clone the repository expecting to pull model weights will receive empty pointer files rather than actual model data.

For anyone building a pipeline that requires ONNX models, the Hugging Face organization at huggingface.co/onnxmodelzoo is now the correct starting point. Hugging Face provides model cards, download links, and community discussion for each model in the collection.

INT8 Quantized Models and the Neural Compressor Pipeline

The repository includes INT8 quantized versions of several models, generated by Intel Neural Compressor. Intel Neural Compressor is an open-source Python library that supports accuracy-driven quantization strategies for ONNX models. It implements both dynamic and static quantization and can represent quantized ONNX models in two ways: operator-oriented and tensor-oriented (QDQ).

The quantized models in the collection are relevant for deployment scenarios where memory and latency matter more than full float32 precision. A quantized model typically occupies less memory and runs faster on compatible hardware, at some cost to accuracy that the Neural Compressor tooling is designed to minimize through tuning.

For developers who want to replicate the quantization pipeline rather than use the pre-built models, the Neural Compressor project documents its approach in its own repository. The ONNX Model Zoo's README points to that documentation for implementation details.

Limitations of the Archived Collection

Deprecation is the most significant limitation. The repository's own notice describes it as preserved for historical purposes only. No new models are being added, no validation runs are being performed, and no accuracy updates are planned. Developers should treat any accuracy numbers cited in the repository as historical benchmarks tied to the model and dataset versions that existed when the entry was written.

A second limitation is the scope of what was validated. The README distinguishes between models the team validated for accuracy and a broader list that was being expanded at the time of deprecation. Not all categories reached the same level of coverage. Generative AI and Graph Machine Learning were listed as expansion areas, which means their model counts and validation depth are likely thinner than the established computer vision and NLP categories.

The repository uses Git LFS for model storage. For historical research involving the metadata and directory structure, a standard clone works. For anything requiring the actual model weights, the Hugging Face mirror is the only option.

ONNX Model Zoo vs. Hugging Face Hub

The ONNX Model Zoo was purpose-built for ONNX format models. It imposed a validation gate: models in the validated directory were checked for accuracy before inclusion. Hugging Face Hub is a general-purpose model repository that accepts models in many formats from any contributor, including ONNX. Hugging Face does not apply the same centralized validation standard; quality and accuracy documentation vary by contributor.

The trade-off is breadth versus curation. The ONNX Model Zoo had a smaller, vetted set. Hugging Face has a vastly larger collection with variable quality signals. For developers who need a specific, historically validated ONNX model, the onnxmodelzoo organization on Hugging Face is the right destination. For developers who need the widest possible model selection in ONNX format, Hugging Face's full catalog offers more options, though each model requires independent evaluation.

The ONNX community has signaled that Hugging Face is the preferred ongoing home for model sharing. The ONNX Model Zoo repository's deprecation notice describes the transition as the natural evolution of the ecosystem. For developers who specifically need the validated model list as a reference, the repository README remains accessible and the table of validated models with their research paper citations still constitutes a useful index of what was considered state-of-the-art at the time the collection was active. The Apache-2.0 license applies to the repository's code and documentation.

Editorial conclusion

The ONNX Model Zoo no longer serves as an active model source. Its LFS files were shut down on July 1, 2025, and the repository is preserved for historical reference only. Developers who need pre-trained ONNX models should go to https://huggingface.co/onnxmodelzoo directly. The repository remains useful for browsing validated model lists, understanding the categories that were covered, and reviewing the INT8 quantization approach for those who need to replicate historical workflows.

Frequently asked questions

Official sources

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