Guidance on redistributing Gemma ONNX weights for a Windows Store App

[levmin]

I am developing a C# based Windows Store application (WinUI 3) that will enable users to download and run Gemma models locally using the Microsoft.ML.OnnxRuntime.GenAI library and DirectML.

The Challenge: To provide a seamless "consumer-grade" experience, I want to avoid requiring end-users to create Hugging Face accounts or generate API tokens. My goal is to allow users to click a "Download" button, view a progress bar, and begin using the model immediately. To comply with the Gemma Terms of Use, the app will display the mandatory Google Gemma EULA and require a click-through "Agree" from the user before the download starts.

Since the official Gemma ONNX weights are currently gated on Hugging Face, a direct unauthenticated download is not possible.

Questions for the Team:

Does Microsoft provide a "blessed" or ungated public endpoint for Gemma ONNX weights specifically intended for Windows-native apps?

If not, does the ONNX Runtime team have specific guidance or "best practices" for developers redistributing these optimized weights to ensure full compliance with the Google/Microsoft partnership?

Are there any technical "gotchas" regarding the GenAI library when loading models from a custom local path rather than the standard Hugging Face folder structure?

I appreciate any guidance you can provide to help make local AI more accessible on the Windows ecosystem.

[kunal-vaishnavi]

Does Microsoft provide a "blessed" or ungated public endpoint for Gemma ONNX weights specifically intended for Windows-native apps?

You can use the official, ungated Gemma-3 vision ONNX models that we have uploaded here. For the text-only Gemma-3 1B model, you can generate it using ONNX Runtime GenAI's model builder. For the Gemma-3 vision models, you can follow this guide to generate your own models.

Are there any technical "gotchas" regarding the GenAI library when loading models from a custom local path rather than the standard Hugging Face folder structure?

A local path is fine to use as long as you have the necessary files in the folder provided (ONNX models and weights, genai_config.json, tokenizer.json, tokenizer_config.json). For Gemma-3 vision, the subfolders in the Hugging Face repo contain all of the necessary files. For Gemma-3 1B, the model builder will generate all of the necessary files in an output folder for you.