ONNX Model Zoo is now a museum piece

What it does

The ONNX Model Zoo was a curated collection of pre-trained deep learning models—ResNet, MobileNet, YOLO variants, BERT-style NLP models, and others—exported into the ONNX interchange format. The idea was simple: give developers framework-agnostic model files they could run anywhere ONNX was supported, without wrestling with PyTorch or TensorFlow directly.

The interesting bit

The README opens with a deprecation notice that is almost startlingly candid. The maintainers explicitly state that “much of the novel model sharing has successfully transitioned to Hugging Face,” and that this repository survives only for historical purposes. Git LFS downloads die July 1st, 2025; the actual models have already migrated to https://huggingface.co/onnxmodelzoo. It is rare to see a project this large—9,600+ stars—gracefully admit obsolescence rather than linger in zombie maintenance mode.

Key highlights

• Coverage spans vision (classification, detection, segmentation), NLP, generative AI, graph ML, and speech/audio models
• Models sourced from timm, torchvision, transformers, and PyTorch Hub, then exported via the TurnkeyML toolchain
• INT8 quantized variants available, produced with Intel Neural Compressor
• Some models include Hugging Face Spaces demos for live testing
• Repository uses Git LFS for the actual model binaries (which is why the LFS shutdown matters)

Caveats

• The deprecation notice is unambiguous: this is not a soft transition, it is a termination of service
• No clarity on whether validation pipelines or export tooling will continue independently

Verdict

Worth a bookmark if you need to trace the lineage of a specific ONNX export, or if you are maintaining legacy pipelines that reference these exact file hashes. Everyone else should go directly to Hugging Face. The model zoo concept was useful; the execution was honest enough to die with dignity.