A developer named Simon Willison just got a two-hundred-million-parameter AI image-inpainting model running in a web browser. Normally, the Moebius model requires Python, PyTorch, and NVIDIA graphics hardware. Willison converted it to ONNX format and used WebGPU—a new web standard for GPU acceleration—to run it directly in your browser. Inpainting is the AI task where you mark parts of an image to remove, and the model imagines what should fill the space. You can try it now at simonw dot github dot io slash moebius dash web. Here's what's clever: Willison did this as a side project while waiting for Claude Code—an AI coding agent—to finish refactoring his main Datasette work. So while one AI was building a UI for his primary project, another was porting a model in a separate terminal. The result: one point two four gigabytes of ONNX model weights, published to Hugging Face, so anyone can build on it. It's a useful example of how AI models are becoming more accessible to web developers without requiring GPUs locally.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton