Find or train models for Windows ML

Windows ML works with ONNX format models, since Windows ML is the Windows-supported distribution of the ONNX Runtime and hardware-specific execution providers. This means you can use millions of existing pre-trained models from various sources, or train your own models. This guide covers where to find, convert, or train ONNX models.

To learn more about Windows ML, see What is Windows ML.

You can also choose from dozens of ready-to-use AI models and APIs in Microsoft Foundry on Windows, which run via Windows ML. See Use local AI with Microsoft Foundry on Windows to learn more.

Option 1: Use models from WinML CLI

With Windows ML CLI, you can start from a verified catalog of over 100 supported models and prepare them for Windows ML. The CLI takes a model from Hugging Face or your own pipeline, then converts, analyzes, optimizes, and compiles it into a portable artifact for Windows hardware.

To browse the available models, see Supported models for Windows ML CLI. To learn more about the tool, see Windows ML CLI.

Option 2: Use other existing ONNX models

Hugging Face hosts thousands of ONNX models that you can use with Windows ML. You can find ONNX models by:

1. Browsing the Hugging Face Model Hub
2. Filtering by "ONNX" in the library filter

You will need to find a model that is compatible with the ONNX Runtime version included in the version of Windows ML you are using. See ONNX Runtime versions shipped in Windows ML to find out what version of ONNX Runtime you are using with Windows ML.

Option 3: Convert existing models to ONNX format

Models from PyTorch, TensorFlow, or other frameworks can be converted to ONNX format and used with Windows ML.

Hugging Face hosts millions of models that you can convert and use with Windows ML.

You will need to convert the model to run with the ONNX Runtime version included in the version of Windows ML you are using. See ONNX Runtime versions shipped in Windows ML to find out what version of ONNX Runtime you are using with Windows ML.

To convert a model to ONNX format, see framework-specific documentation, for example:

• Convert PyTorch models tutorial
• Convert TensorFlow models tutorial

Option 4: Fine-tune existing models

Many models on Hugging Face or other sources can be fine-tuned (following instructions on the model cards on Hugging Face). You can then subsequently convert the fine-tuned model to ONNX following the instructions in Option 3 above.

A popular way to fine-tune models is to use the olive finetune command. See the Olive documentation to learn more about using Olive.

Option 5: Train models

If you need a model for a specific task and can't find an existing model, you can train your own in PyTorch, TensorFlow, or other frameworks.

Once you've trained your model, follow the instructions in Option 3 above to convert your model to ONNX format.

Next steps

Once you have an ONNX model, you can run it with Windows ML on your target devices.

• Install Windows ML EPs - Download and install execution providers in Windows ML
• Run ONNX models - Learn how to run inference with Windows ML

Other solutions

As part of Microsoft Foundry on Windows, you can also choose from dozens of ready-to-use AI models and APIs, which run via Windows ML. See Use local AI with Microsoft Foundry on Windows to learn more.